What the researchers did

Research on reviews stops at the till. We know a great deal about how reviews move purchases and almost nothing about what they do afterwards, because of an unspoken assumption that once you have the thing in your hands, your own senses take over.

Dena Yadin, Nira Munichor, Elanor Williams and Inbal Stockheim went looking for the afterlife. Their setup is the same every time and it is the right one: everybody gets the identical product, and the only thing that differs is which reviews they read first. Any difference in how much they enjoy it therefore belongs to the reviews and not to the thing itself.

Across the main studies people watched music clips, read a book excerpt, listened to a holiday song and watched part of a ventriloquist act. The supplementary studies add earbuds, a virtual tour and an origami kit. The reviews were real, written by earlier participants who had genuinely rated the same material, which matters: these are not researcher-written caricatures of a bad review.

Their survey work also establishes that this is not a rare situation. Most people read reviews after buying, after being given something as a gift, and during or after using a product.

Main studies

Five, plus six supplementary replications and several more reported in the appendix

People

4,291 across the five main studies alone, recruited through Prolific

Design

Everyone experiences the identical product; only the reviews they read beforehand change

Reviews used

Real reviews written by earlier participants, shown with a two-star or four-star average

Measured

Star rating, a five-item enjoyment index, what people recalled, what they wrote, and what they would pay

Bad reviews work. Good ones do not.

Study one put 719 people in front of one of two music clips, one pretested as better and one as worse, after showing them five negative reviews, five positive reviews, or nothing at all.

Those who saw the negative set came away liking the clip less than the group shown no reviews at all. The positive set landed its readers exactly where that no-review group sat, with a gap nowhere near significance. That lopsidedness is the finding the whole paper is built on.

It held for the good clip and the bad clip alike, with no interaction between them, and that is what rules out the comfortable explanation. If this were about expectations being met or missed you would expect bad reviews attached to a good product to produce a pleasant surprise. They did not. The experience bends towards the review rather than away from it.

Six supplementary studies then tried to break it and failed. Framing the product as useful or as fun, letting people choose it or assigning it, matching the positive and negative reviews word for word so only the valence differed, using one-star and five-star extremes, and inserting delays of minutes and then of a full day. Every one reproduced the same shape.

It is the experience, not the write-up

Here is the obvious objection, and the authors take it seriously enough to build a study around it. Maybe nobody's experience changed at all. Maybe people simply adjust what they say to match what everyone else said.

So study two gave 999 people the same book excerpt and varied only when the bad reviews appeared: before choosing a book, after choosing but before reading, halfway through, or after finishing and just before giving a rating. That last group cannot have had their reading experience altered, since the reading was over. Any drop in their ratings is pure reporting.

Their ratings did drop, which tells you conformity is real. But they dropped less than everyone else's. The people who read the reviews before or during their reading came out at 2.60 to 2.68, while the people who read them at the end came out at 2.87. Reporting explains part of this and cannot explain all of it.

One more thing worth lifting out of that study. Reading the reviews halfway through was as damaging as reading them at the start. The lens fits at any point you put it on.

One is enough

In the real world reviews are mixed, and the practical question is whether one bad one among several good ones actually matters. Study three, with 597 people, answered it cleanly by holding the set size at four and changing only the fourth review.

Swapping one of four positive reviews for a negative one pulled ratings down measurably. Swapping one of four negative reviews for a positive one did nothing whatsoever. The two conditions came out a hundredth of a point apart.

That is negativity dominance in its purest form, and it is the result I would put in front of anyone who believes a wall of five-star reviews provides cover. It does not. The one-star review in the middle of it is still doing its work on the person who buys.

What is actually going on

Study four explains the mechanism, and the measure is elegantly simple. After listening to a holiday song, 376 people were asked to quickly list five things they remembered about it, then to mark each one as something they liked or disliked.

Everybody listed five things. What changed was how many of them were complaints. The people who had read negative reviews filled more of their five slots with things they disliked, and that proportion statistically carried the effect through to how much they said they enjoyed the song.

So the reviews are not planting an opinion. They are redirecting attention. You go in primed to notice the thin vocal and the cheap drum machine, you duly notice them, and a song made of the same notes as everyone else's comes out worse. The authors call it a negative lens, and the lens is made of what you are looking for.

There is a second-order effect here that marketers should sit with. Everyone in these studies wrote a review of their own afterwards, and the ones primed by complaints wrote harsher ones. The thing feeds itself.

The fix costs one sentence

Study five is the practical one, and the intervention is almost insultingly cheap. 1,600 people read either positive or negative reviews before hearing the song. Half of them were additionally told, in advance, that reviews can colour an experience and asked to reach their own verdict.

For the half who had seen the complaints, that single prompt recovered a good share of the loss. Among the people who read positive reviews it changed nothing at all, which is what makes it safe to deploy: it cannot cost you anything on the products whose reviews are already good.

Timing is the catch. The prompt has to arrive before the impression forms, not after, because negativity is famously hard to dislodge once it has taken hold. In practice that means the gap between the order and the first use, which for most businesses is the confirmation email, the packaging insert or the first screen of an app.

Going past what was tested, the implication for review management is broader than it looks. The usual cost-benefit on replying to a bad review is about the next prospect reading the thread. This paper says there is a second bill arriving: the people who read it, bought anyway, and will now enjoy your product less and write a worse review because of it. That changes which reviews are worth answering and how quickly.

What to expect

• Good reviews will not do the opposite job. Collecting more five-star reviews protects the sale, but it buys you nothing in the experience itself. Only the bad ones carry over.
• The harm reaches people who never bought. Anyone who reads the reviews and then uses the product, a guest, a gift recipient, a colleague, carries the same lens in.
• It compounds through the next review. Readers of negative reviews wrote more negative reviews themselves, which is how one complaint becomes a trend line.
• The prompt is free to try and hard to get wrong. It helped where the reviews were bad and did nothing where they were good, so the downside in the tests was zero.

Where it does not work

• Short experiences, not long ownership. Music clips, a book excerpt, a song, part of a show, earbuds and an origami kit. Nothing here runs over the months of living with a purchase.
• Mostly self-reported enjoyment. The willingness-to-pay study is the one behavioural outcome. The rest is what people say about the experience, however carefully the design separates that from the experience.
• Online panels throughout. Prolific participants given a product to try, not shoppers spending their own money in a shop.
• The fix was tested in the lab, in one line. Nobody has yet run it as a real packaging insert or confirmation email, where tone and placement will matter.

Questions marketers ask

Also new this week

What the researchers did

Anastasia Nanni of Aalto, Aulona Ulqinaku of LIUC and Verdiana Giannetti of Leeds picked a question the literature had skipped. Research on scandals assumes one culprit: either the brand did something wrong, or the celebrity did. But an influencer campaign is made jointly, the brand hands over a good deal of control, and when the result deceives people nobody can say cleanly whose fault it was. So who does the public punish?

Their starting case is a good one. For Christmas 2022 the Italian influencer Chiara Ferragni worked with the confectioner Balocco on a special edition pandoro. The packaging and the posts both said a share of the money would go to the Regina Margherita children's hospital in Turin. In December 2023 investigators found that a €50,000 donation had been made up front, before the cake went on sale, and that not a cent of the money from actual sales ever reached the hospital. Both parties were fined for misleading advertising.

The authors took 80,314 consumer tweets from before the collaboration and after the fines, then added 20,120 tweets from journalists and news outlets, two surveys of practising journalists and editors, and six experiments using invented brands and invented influencers. They also took the unusual step of publishing a theory they had to abandon: they expected blame to follow power, and it did not.

The case

Chiara Ferragni and Balocco, a Christmas pandoro sold on a charity claim, fined in December 2023

Consumer data

80,314 tweets, comparing January to October 2022 against December 2023 to July 2024

Press data

20,120 tweets from accounts verified as journalists or news outlets

Journalists surveyed

200 working journalists and editors in the US and UK, plus 225 more in a later study

Experiments

Six, using fictional influencers and brands, with consumers recruited in Britain and America

The influencer takes it

The design here is simple and clean. Tweets about Ferragni are the treatment group, tweets about Balocco are the control, and the comparison is how sentiment in each group changed from before the collaboration to after the fines.

Sentiment fell for both, as you would expect once a fraud is in the news. It fell considerably further for the influencer. That result survived clustering the errors by user, adding controls for gender, country and follower count, adding user fixed effects, and matching treated and control users. Under the strictest specification the gap was larger, not smaller.

The sheer imbalance in what people chose to talk about tells its own story, and it is in the takeaways above. Nobody was tweeting about the biscuit company.

Not fame, and not money

The obvious explanation is attribution: whoever looks most in charge gets the blame. The authors tested the two versions of that and both failed.

The first was prominence. They built a fictional charity-fraud scenario and swapped the cast: a hugely famous influencer or a small one, paired with a huge beauty retailer or a small brand. Four combinations, 401 people. Prominence made no difference to attitudes toward either party, on its own or in combination.

The second was money, on the theory that the party with deeper pockets carries more responsibility. In a second experiment the influencer was either a multimillionaire paired with a tiny startup, or nearly broke and paired with a market leader. Attitudes toward the influencer did not shift at all. Attitudes toward the company shifted slightly, but in a way that made no sense as an explanation, since it moved the same amount whichever party had the money.

This is the part of the paper I respect most. Having lost both of their theories, the authors say so in print and go back to the data rather than quietly reframing the paper around whatever turned up significant.

The press decides first

What they had overlooked is that almost nobody learns about a scandal from the parties involved. They learn about it from the press. And the question is not how much coverage there is, which tracks fame, but who the coverage is about.

In the journalist tweets the split was not close. Having established that, the authors checked whether it was a quirk of one Italian story by giving a fresh panel of reporters and editors an invented transgression by an invented influencer and an invented company, with no follower counts and a deliberately gender-neutral name, and asking them to write a headline and say who they would cover. Most chose the influencer. Their headlines skewed negative, and the negative ones came disproportionately from the journalists who had chosen the influencer.

Then they asked those journalists to explain themselves, and clustered the answers. Most of the reasoning came down to the audience relating to a person rather than an organisation. Some was frank about traffic: a person gives the story a face. One cluster the authors labelled gloating, and the quotes in it are unsparing about how much more fun it is to watch a named human fall than a faceless company.

One more result closes the loop. Taking the daily average sentiment of journalists and consumers, journalists' sentiment on one day predicted consumers' sentiment the next. The press was not reporting the mood. It was setting it.

Change one name in the headline

To prove that coverage is doing the work rather than merely accompanying it, the authors wrote a set of headlines about an invented scandal and showed them to 298 people. Every headline was identical in wording, tone and severity. The only thing that changed between conditions was which party was named in them.

That single edit moved how people judged the influencer, and left their judgement of the brand statistically untouched. Being named is what costs you, and it costs an influencer in a way it does not cost a brand.

Why only the influencer? A follow-up with 301 people put four explanations in the same model, each drawn from what the journalists had said: that consumers see through journalists' self-interest, that influencers are held more accountable, that people enjoy the downfall, and that people identify with the influencer. Only identification carried the effect. The measure is a neat one, a set of overlapping circles where you pick how much of you overlaps with them. Media criticism shrank that overlap for the influencer, and the smaller the overlap, the harsher the attitude.

So the mechanism is self-protective rather than moral. You liked her, you saw a bit of yourself in her, and now she is in the paper for stealing from a children's hospital. The cheapest way to feel decent again is to decide you were never really a fan.

The way back is the interview you least want

The final pair of studies is the practical payoff, and the answer is counterintuitive enough to be worth the whole paper.

Three responses were tested, differing only in how much control the influencer keeps. Post to your own followers and you keep all of it. Sit with a sympathetic journalist and you keep some. Sit with a journalist who has been hard on you and you keep almost none.

Shown each version, 225 journalists were markedly less willing to tear into the influencer who had faced the hostile interviewer, and markedly less generous to the one who had posted to followers. The numbers are in the takeaways. Then 302 consumers saw coverage of each response, and their attitudes followed the same order.

The lesson runs against every instinct in crisis communications, where the standing advice is to control the message, speak directly to your audience, and avoid anyone looking for a scalp. On this evidence, controlling the message is exactly what makes it worthless. Submitting to scrutiny is the costly signal, and a costly signal is the only kind that gets believed.

A note for brands, since most readers here sit on that side. The finding that the press shields you is not a licence. It is a description of a risk you are handing to someone else, and the authors are blunt that it should push companies toward better vetting, clearer contracts and shared accountability rather than relief. My own view, going past what was tested: if your partner takes the beating for a campaign your legal team approved, you will not get a third campaign out of anyone good.

What to expect

• The brand gets off lighter than it deserves. Coverage concentrates on the person, and feeling toward the company shifts very little. Plan for the asymmetry rather than being relieved by it.
• Your partner may not come back. An influencer who absorbs the damage for a campaign you signed off has every reason to decline the next one, and to tell others why.
• A friendly interview buys less than you think. It sat between the hostile interview and the own-channel post on both journalist and consumer measures, closer to doing nothing.
• Legal clearance is not reputational clearance. Both parties in the case had approved the wording. The fines and the backlash arrived anyway.

Where it does not work

• One real case, and an Italian one. A beloved national figure, a Christmas product and a children's hospital. The experiments used invented actors and English-speaking panels, but the field evidence rests on a single story.
• The field data stays correlational. The before-and-after design is well defended and the controls are thorough, but nobody randomised who got investigated.
• Attitudes, not sales. Every experiment measures how people feel about the parties. None follows the money, the unfollows or the next quarter's revenue.
• It is about shared blame specifically. Where fault is unambiguous, the older research on brand scandals and celebrity scandals is the better guide.

Questions marketers ask

Also new this week

What the researchers did

In September 2023 TikTok became the first big platform in the United States to put an AI label on posts. That gave Stephan Carney, Ignacio Riveros and Stephanie Tully, all at USC Marshall, a natural before and after to work with, and they used it properly.

First the field data: every post from roughly ten thousand randomly chosen American TikTokers for three months before the policy and six months after, with the engagement each one collected and whether it carried the label. Then, because that can only tell you what happened and not why, eight preregistered experiments, a content analysis of four hundred videos stripped of their interface, and a linguistic sweep of nearly two million comments.

Their hypothesis was unusual. Most people assume an AI label works by raising suspicion: this might be fake, this might be low quality, I do not trust it. The authors proposed something else, borrowed from the research on why people like each other. We judge a relationship partly by how much effort the other person appears to put in. A viewer and a creator have exactly that kind of one-sided bond, and a label announcing AI is read as a creator who did not bother.

Field data

1,135,817 TikTok posts from 8,650 creators, spanning the policy launch

Comments analysed

About 1.9 million, across 56,720 posts

Experiments

Eight preregistered studies, 3,396 participants in all

Content checked

400 videos transcribed, scored for brightness, sharpness, motion and theme, with the interface stripped off

Outcome measured

Likes, comments and shares on the platform; likelihood of tapping like in the experiments

What happened on TikTok

Comparing posts by the same creator in the same month, and holding views constant, posts carrying the label collected less of everything. Fewer likes, and less combined engagement once comments and shares were added in.

Holding views constant matters more than it sounds. Labelled posts were also shown to far fewer people, which could be the algorithm or could be the creator posting different sorts of thing. By conditioning on views the authors are asking a narrower and more useful question: of the people who did see this, how many tapped like. The answer is fewer.

One finding should worry anyone running influencer budgets. The effect did not belong to small accounts. The authors split the data by follower count, into accounts under a thousand followers, accounts up to ten thousand, and accounts above that, then split it again by how many views a creator had been getting before the policy existed. The penalty turned up across almost every group, and there was no significant interaction with audience size. Whatever this is, it happens to the mid-tier creators brands actually hire.

It is worth being honest about the raw numbers behind all this. In the post-policy window the dataset held 819,227 unlabelled posts and only 2,749 labelled ones. The label was rare, and the people applying it sometimes chose to.

It is not the quality, and it is not AI

Three separate attempts were made to kill the obvious explanations, and all three failed to save them.

The first took 200 pairs of videos from the same creators, one labelled and one not, stripped off the TikTok interface and ran them through transcription, language analysis and computer vision: brightness, colourfulness, sharpness, motion, scene cuts, topic. Feeding all of that into the model did not make the label's effect go away.

The second showed those same videos, without any label visible, to 697 students who use TikTok heavily, for 5,277 ratings. Perceived quality came out at 3.04 for the unlabelled videos and 2.98 for the labelled ones, a difference so small it barely registers, and once it was added to the model it predicted nothing while the label kept predicting. Creativity ratings did not differ at all. Here is the detail I find most telling: viewers guessed AI had been used in 36% of the labelled videos and 8% of the unlabelled ones, and the videos people believed were AI did not lose engagement. Believing is not the same as being told.

The third read the comments. If people thought labelled posts were worse, the comments should be more negative. They were slightly less negative. If people were worried about being deceived, there should be more language about risk or right and wrong. There was not. What did shift was the balance of pronouns: comments on labelled posts used relatively fewer personal words like I, you and she, and more impersonal ones like it and that. People were talking about the content rather than to the person who made it.

Finally, the AI aversion story. In one experiment 56% of participants said they liked AI and fewer than a quarter said they disliked it, the effect appeared anyway, and neither their attitude to AI nor how much they used it changed the size of it.

What the label actually says

So what is left? Effort.

Shown the same video with and without a label, people rated the creator as having put in substantially less work when the label was there. Their sense of connection to that creator fell with it, from 4.01 to 3.80 overall, and from 4.01 to 3.43 among the people who actually noticed the label. A formal test of the chain, label to perceived effort to connection to engagement, came out significant, and once effort and connection were in the model the label had nothing left to explain on its own.

That is the whole mechanism, and it is a social one rather than a technological one. Nobody is deciding the video is fake. They are deciding the person behind it did not show up, and a like is a small social act you perform for someone you feel something for.

Owning up does not help

The intuitive fix is honesty with context: say it yourself, in your own voice, and explain that AI only did something trivial. One experiment tested exactly that, with 565 people and a photograph of a Japanese breakfast by a window, where AI had added falling snow. The caption version read like a creator being cheerful about it: check out the snow effect I added.

It made no difference whatsoever. Disclosing it yourself scored the same as being labelled by the platform, and having both was no worse than either. Every version carrying any admission of AI landed well below the version with none.

Read that against the mechanism and it makes sense. The caption was not reassuring anyone about deception, because deception was never the issue. It was confirming the thing that costs you: a creator who reached for a tool instead of doing it.

The wording is the whole lever

Here is the finding worth the price of admission. In the last experiment, 198 people saw the same posts with three different labels: none, a generic AI disclosure, and one naming a specific tool, Photoshop neural filters. Photoshop is AI. It can generate and alter an image as thoroughly as anything else. But a pretest confirmed people read it as requiring more work.

The generic label cost engagement. The specific one cost nothing measurable at all, landing level with no label. Sense of connection to the creator followed the same shape. And because the actual content was shuffled between the three label conditions, the content itself had no effect in the model. The words under the post were doing all of it.

That hands platforms a design decision rather than a trade-off. YouTube Shorts has already started naming tools in its disclosures rather than announcing AI in the abstract. On this evidence that choice keeps the transparency and drops most of the penalty.

There is also a warning in here for policy, and the authors make it plainly. If a label reliably costs creators reach and money, creators will find ways around it, which defeats the point of having one. A disclosure regime that punishes honesty does not produce honesty. My own reading, beyond what they tested: the same logic should apply to any provenance label a brand is now being asked to carry, from AI notices to sponsorship tags. What the label says about the person behind the work is doing more than the fact of the label.

What to expect

• A steady tax, not a collapse. The hit to engagement at a given reach is a few percent. It matters because it compounds across every post and feeds the algorithm, not because any one post dies.
• Reach is a separate problem. In the raw data, labelled posts reached a fraction of the audience before anyone chose whether to like them. The study cannot say why, but budget for both effects.
• Half your audience never sees the label. That caps the damage today. It also means the penalty grows as platforms make disclosures more prominent.
• Specific wording is free to implement. Changing a generic label to a named tool costs nothing, needs no new content, and in this test recovered the whole difference.

Where it does not work

• One platform, one policy, nine months. TikTok's first labelling window, in the United States. Other platforms place and word their labels differently, and Instagram buries its one in the description.
• Labelled posts were rare and partly self-selected. Creators applied the label themselves when they used outside tools, so the labelled group is not a random sample of AI-assisted content.
• The experiments measured intention. Participants said how likely they would be to tap like. Only the TikTok data records what people actually did.
• Norms are moving underneath this. The authors note the effect could shrink as AI use becomes ordinary, or grow as the tools get easier and the perceived effort falls further.

Questions marketers ask

Also new this week

What the researchers did

Almost nobody gets to run this experiment, because almost nobody will let a researcher switch off a working ad channel. Jeffrey Larson and Jeffrey Dotson, both at Brigham Young University, persuaded a large direct-to-consumer mattress company to let them do exactly that.

The setup is the kind of thing an agency talks about and never builds. They split the United States by Nielsen media market, 210 of them, and assigned each market to one of 19 budget recipes. A recipe set the spend separately for four things: Google ads to strangers, Google ads to previous site visitors, Facebook ads to strangers, Facebook ads to previous visitors. Some markets got nothing in a channel, some got half the usual budget, some the usual amount, some half again as much. Then everything ran untouched for nearly two months while the orders came in.

In the industry's language, ads to strangers are prospecting and ads to previous visitors are retargeting. The company, like the three agency heads the authors interviewed, was certain retargeting was the better buy. It was certain enough that it refused to let the researchers raise the prospecting budget above normal, on the grounds that spending more there would dilute the targeting.

Design

19 budget recipes assigned across 210 US media markets, about 10 markets per recipe

Budget moved

$5.7 million of display spend over nearly two months, against roughly $58 million of sales

Four channels

Google and Facebook, each split in two: ads to strangers, ads to people who had already been to the site

Levels tested

Nothing, half, normal, and half again, so saturation could be seen as well as lift

Category

Direct-to-consumer mattresses, average online order around $1,000

What the platforms were reporting

Start with the numbers the company was actually managing to. During the experiment itself, the two platforms reported their own returns on every dollar: Facebook claimed its retargeting ads were returning more than three times what its prospecting ads returned, and Google reported retargeting ahead too.

Those figures had shaped years of decisions. They are also the reason the industry treats the combination of prospecting and retargeting as a one-two punch, with retargeting doing the finishing. The belief is not unreasonable on its face. Someone who has been to your website has told you something real about their interest, and nobody in the prospecting pool has told you anything at all.

There was a quieter warning sign in the data, and the company had not read it. In the markets told to spend half again as much on retargeting, the platforms could not actually spend the money. The pool of previous visitors simply did not generate enough impressions to absorb it. A budget that cannot be spent is a budget that was already too big.

What happened to the orders

Now the part the platforms cannot see. Because budgets were set by geography and the company had never run regional campaigns before, the markets work as controls for each other, and the order counts can simply be compared.

On prospecting, more money meant more orders, cleanly and in both channels. On retargeting, nothing. Not a smaller effect, nothing measurable at all, in either channel, at any level.

The detail that kills the obvious objection is the step from nothing to half. If retargeting were merely oversold, if the company were buying the tenth impression when the second would have done, then cutting the budget in half should still have hurt less than cutting it entirely, and the markets at half spend should sit somewhere in between. They do not. Going from no retargeting at all to half the usual retargeting produced no improvement that came anywhere near significance.

Two further results are worth having. Advertising in no channel changed how much people spent once they decided to buy, only whether they bought. And the experiment was built to detect synergy between prospecting and retargeting, the idea that prospecting fills the pool that retargeting then converts. No such interaction showed up, though the authors are careful to say the test was underpowered even at this budget.

Why the dashboard was wrong

The authors' explanation is mechanical, and once you see it you cannot unsee it. Someone sees a prospecting ad, clicks, lands on the website, and does not buy, because nobody buys a mattress in an afternoon. For the next fortnight they are shown retargeting ads constantly, and the ads work the way a shopfront works: they recognise the product they are already thinking about, so sooner or later they click one. Then they buy. The platform looks at that final click and awards itself the sale.

Retargeting, on this reading, is not causing purchases. It is standing next to them. And the longer the purchase cycle, the more reliably it gets to stand there, which means the bias is worst in precisely the categories where the stakes are highest.

There is a second reason, less about measurement and more about what an ad is for. If advertising works mostly by telling people things they did not know, then the retargeting pool is the worst audience you have, because they already went to your website and read it. The prospecting pool is the only group left who can still learn something. That logic also puts a question mark over any channel that targets people the platform has already judged likely to buy. The authors name Google's Performance Max as the obvious candidate for the same flattering arithmetic.

What it costs, and what to do

Turn the elasticities into money and the picture is uncomfortable for the budget as it stood. One extra order cost far more through Facebook prospecting than through Google prospecting, and during the experiment Google prospecting was the smallest of the four budgets at $0.76 million.

Run the same arithmetic on the other side and it says Google prospecting, which had been running at $1.79 million before the experiment, should have been somewhere near $4.14 million. More than double, into the channel that was getting the least.

The authors are realistic about most firms not being able to run a test like this, so their fallback recommendation is a single setting: frequency caps. Cap the retargeting and the spend falls without the returns falling, since the returns were not there. Cap the prospecting and the same budget reaches more people instead of the same people more often, which is what you want if information rather than persuasion is doing the work.

One finding that the authors flag and I would underline: Google prospecting ads also lifted the company's Amazon sales, where it had no ads at all and sold only its cheap accessories. People saw an ad, went looking, and bought where they already had an account. Nothing in any dashboard would ever have shown that.

What to expect

• A smaller bill before you see anything else. Capping retargeting cuts spend immediately. The orders are the slow part, so give it a month before judging.
• Your reports will look worse while your business does not. Credit moves away from the channel that was claiming it, so reported returns fall even if total orders hold or rise.
• Prospecting pays off on a delay. Here the typical gap between first visit and purchase was a fortnight, so a week of data will show you the cost and none of the benefit.
• It changes who buys, not how much they spend. No channel in this experiment shifted order value, so do not expect a bigger basket out of a budget move.

Where it does not work

• One advertiser, one category. A high-priced mattress bought after weeks of thinking. Cheap, impulsive, repeat purchases may behave differently, and the authors say so.
• Collected in 2019. Before Apple's tracking prompt and the wider privacy changes, which have since made retargeting pools smaller and blunter rather than better.
• The budget only went down. The client refused to let prospecting rise above normal, so the case for spending much more there is an extrapolation from the measured elasticity, not a tested result.
• No synergy found is not no synergy. The test for prospecting feeding retargeting had standard errors wide enough that only an enormous effect would have registered.

Questions marketers ask

Also new this week

What the researchers did

Plenty of research asks what a photo does to a review once it is posted. Almost none asks where the photo came from. Jiani Xue, writing as an independent researcher, and Shiri Melumad of the Wharton School went after the thirty seconds before the shutter.

Their argument is simple enough to state in a line. Taking a decent picture is work, you have to frame it, clear the clutter, find the light, and people are willing to do that work for something they like and not for something they resent. So the pictures attached to praise come out better than the pictures attached to complaints, and because better pictures read as more useful, the whole review inherits the difference.

They tested it six ways: a large online experiment where people photographed their own clothes, a lab study where the product was held visually identical and the participants were filmed without knowing it, two studies where strangers rated the resulting pictures, and finally two very large piles of real reviews from Amazon and Yelp. Picture quality was never judged by the authors. It came from independent human raters, from MUSIQ, a deep learning scorer trained on smartphone photography, and from GPT-4o prompted to react like an ordinary shopper.

Studies

Six: two on effort, two on how observers react, two on real review data

People in the lab and online

4,218 across the controlled studies, from Prolific, CloudResearch and a university lab

Real reviews analysed

348,108 Amazon reviews of 500 products, plus 321,829 Yelp reviews of 186 Los Angeles restaurants

Photos rated

Over 100,000 by algorithm, thousands more by recruited human judges

Three ways to score a picture

Human judges, the MUSIQ deep learning model, and GPT-4o acting as a shopper

One thing you love, one thing you regret buying

The first study asked 1,313 people to do the same task twice. Find a piece of clothing or an accessory you are really happy with, write a review of it, photograph it. Then do it again for something you are really unhappy with. Same person, same phone, same room, so the only thing that moved was how they felt about the object in front of them.

Effort moved with the feeling. Asked how much care they had taken, whether they staged the shot, whether they minded about quality, they came out at 4.65 out of seven for the garment they were pleased with and 4.18 for the garment they regretted. They also took more frames before settling on one, 2.12 against 1.95.

Then the pictures went out to independent judges who had no idea which review each came from. They rated the photos of liked items higher, 3.56 against 3.29. Small numbers, and the same direction turned up in both machine scores. A real gap, running quietly through everyone's camera roll.

The cookie test

This is the study worth retelling at dinner. 144 students were sat at desks with two cookies in front of them, labelled A and B. The cookies looked the same. One had been baked to taste good, the other baked to taste bad. On each desk was a box of props from a fictional bakery: pinecones, dried leaves, dried flowers, decorative paper.

Participants tasted, wrote a review, then photographed each cookie for the review. What they were not told until afterwards was that three cameras were filming the desk. A research assistant later sat with a stopwatch and timed, for each cookie, how long each person had their phone camera open.

The gap was not subtle. Self reported effort came in at 4.10 for the good cookie and 2.88 for the bad one, and the stopwatch backed it up. Judges then called 58.9% of the good cookie photos well composed against 30.1% of the bad cookie photos, which is the part that matters most. The two cookies were visually identical. Nobody can argue the better pictures came out better because the subject was prettier.

Why a better picture is a more useful picture

So far this is a story about photographers. The next two studies turn it around and look at the person scrolling.

347 observers were shown photos of shoes taken for real positive and negative reviews, with the review text above each one. They rated the positive-review photos higher on overall quality, 4.25 against 3.73, and, more interestingly, reported more trust in the person who took them, 4.92 against 4.72. Helpfulness followed the same line, 5.40 against 5.14. Two things are happening at once: a well made picture is easier to take in, and visible effort makes the reviewer look like someone worth listening to.

The obvious objection is that nice shoes photograph nicely. So the authors took twenty photos that judges had already agreed were good, half showing attractive shoes and half unattractive ones, and ran each through PowerPoint to make a degraded twin: blurred, off centre, exposure pushed around. Fresh observers rated one image each for helpfulness.

Degrading the picture cost it a lot of helpfulness. How appealing the shoes were made no difference at all, and it did not interact with image quality either. The craft of the photograph was doing the work, not the product in it.

Nearly 670,000 reviews say the same thing

Then the field data, and here the scale changes completely. Eight years of Amazon reviews covering shoes, bags and hats, of which 24,996 came with at least one product photo. Nineteen years of Yelp reviews covering restaurants across greater Los Angeles, of which 82,472 came with photos.

Both datasets replicated the pattern on every measure. Star ratings and GPT-scored picture quality moved together on Amazon, a correlation of .34, and on Yelp, .21. The MUSIQ scores leaned the same way, 63.23 for positive Amazon reviews against 61.26 for negative ones. Human judges, shown a sample with no idea where each picture came from, agreed.

The last study asks what this does to a ranking. Reviews with photos collect more helpful votes, that much was already known. And negative reviews have always collected more helpful votes than positive ones, which is also old news. The new finding is what happens when you put those two facts together. On Amazon, adding photos to the picture did not just weaken the old negative relationship between star rating and helpfulness. It flipped it. Among reviews with images, the more positive the review, the more helpful votes it got. On Yelp the same interaction appeared but was not strong enough to reverse anything, only to soften it.

What a review page is actually showing you

Put the chain together: a happier customer takes more care, more care makes a better picture, a better picture reads as more helpful and more trustworthy, more helpful votes buy a higher position on the page. Nobody designed that. It falls out of how people feel about holding a camera.

Two consequences follow for anyone running a review page or reading one. The first is a measurement problem. If your platform ranks on helpful votes, it is partly ranking on how motivated your reviewers felt, and that favours the praise. The second is a judgement problem, and it is the one I would put on a wall. A blurry, badly lit, awkwardly cropped photo of a fault is not weak evidence. It is evidence that nobody bothered to dress it up. The polish of a picture tells you about the mood of the person holding the phone, not about whether the product works.

Going beyond what was tested here, the same logic should reach any user content you solicit: unboxing clips, before and after shots, case study photos from clients. The material you get back will be prettiest from the people who needed least convincing. There is also a note for service teams, since Yelp reviewers wrote about three times as many words as Amazon reviewers, 81.74 against 28.05. Where the experience is complicated, people fall back on language, and the photo carries less of the load.

What to expect

• The gallery you build will flatter you. Your best pictures will come from your most satisfied customers, so the page will look better than your average experience actually is.
• Photo requests convert poorly in absolute terms. You are asking for work, not a click. Measure per hundred requests and expect a small number.
• Polish is not proof. A good-looking photo earns trust the facts behind it may not deserve. Train your team to read a scrappy complaint photo as undressed, not as unfounded.
• It will not soften the words. The process behind writing a review runs separately from the one behind photographing it, so a photo request does not improve the text that comes with it.

Where it does not work

• A short list of categories. Clothes, shoes, bags, hats, cookies and restaurant meals. Nothing here tests services with no physical object to photograph.
• Quality was scored, not experienced. Pictures were rated by recruited judges and two algorithms, which is consistent and repeatable but not the same as a shopper glancing at a thumbnail for a second.
• The field patterns are associations. The causal claims rest on the experiments. Amazon and Yelp add scale and realism, with controls, not proof of cause.
• Helpful votes are not sales. Every outcome measured here is a judgement of helpfulness or a vote, never a purchase, a return or a refund.

Questions marketers ask

Also new this week

What the researchers did

A thumbnail is the only part of a video most people ever see. Jiaming Fang, Qiqi Zheng, Miyan Liao, Lintong Han and Lu Liu, working across three Chinese universities, set out to find what separates the ones that get clicked from the ones that do not.

They scraped 22,958 videos from Bilibili, the Chinese platform closest to YouTube, across eight categories from cookery to science, and tracked each one daily for a month. Then they turned the thumbnails over to two trained models. One scored how much was going on in each picture: the number of objects, the colours, the edges, the clutter. The other scored how well the colours and textures in a picture fitted together.

Popularity combined views, comments, likes, shares and saves. To check that the correlations meant something, they then built nine thumbnails of their own, varying the same two qualities, and put them in front of 423 people browsing a mocked-up page.

Thumbnails

22,958 videos tracked daily for 31 days on Bilibili

Categories

Eight, from study and finance to cute pets, comedy and science

Two qualities measured

How much detail a picture holds, and how well its colours agree

Popularity

A combined measure of watching and reacting, with views counted only after five seconds

Experiments

423 people shown nine made-up thumbnails, then 322 more in a second test

Busy is better, up to a point

The first rule is the one people half expect. More going on in a thumbnail earns more clicks, right up until it does not. The curve rises and then falls, and the top of it sits in the middle.

In the experiment, the middling version was clicked far more readily than either extreme, and the margin was wide. A nearly empty thumbnail gives the eye nothing to catch on. A packed one asks for more work than anyone browsing is willing to do, so they scroll on.

This holds in the real data too, across all 22,958 videos, and it holds whether a human chose the frame or the platform picked one automatically. That last point is worth sitting with. The same rule applies to the default frame your platform grabs, which is the thumbnail most small creators and most brands are quietly running with.

With harmony, the middle is the worst place

The second rule is the surprise. Colour harmony does not simply help. Thumbnails where everything matched did well, and so did thumbnails where nothing matched at all. The ones that lost were the half-hearted middle.

Two different things are going on. A harmonious picture is easy and pleasant to take in, so it reads fast. A clashing one is jarring enough to be interesting, which stops the scroll for a different reason. A picture that is neither does neither job.

The authors tested whether the two rules interfere with each other and found that they do not. You can set the amount of detail and the degree of harmony separately, which makes this one of the rare design findings you can act on without a trade-off.

Who this matters most for

The effects were sharpest for two kinds of video. Useful ones, where the viewer wants to learn or decide something, mattered more than entertainment. So did typical ones, meaning videos that look like everything else in their category, where the thumbnail is the only thing doing the separating.

Viewer type mattered too. People who were already interested in the subject reacted more strongly to how busy the picture was. For them a crowded thumbnail is genuinely more work, because they are actually reading it rather than glancing past.

Interestingly, that pattern did not repeat for harmony. Whether a picture hangs together seems to register the same way for everyone, interested or not, which fits the idea that harmony is processed almost instantly while detail has to be worked through.

Beyond the thumbnail

The study is about video, but the situation it describes is everywhere: a grid of competing images, a second of attention each, one of them has to win. Product listings on a marketplace, creative variants in an ad carousel, app icons on a crowded store page, the hero image at the top of an email.

Going past what was actually tested, the two rules translate neatly. Give the eye three or four things to look at rather than one or ten. Then decide whether your picture is going to be the calm one or the loud one in the grid, and commit, because the version that hedges is the version that gets scrolled past.

The caveat worth carrying is cultural. This is Chinese data, and interfaces there run busier than Western ones by habit. The shape of the curves should travel. The exact place the peak sits probably does not.

What to expect

• A click effect, not a watch effect. This changes who starts the video. It does nothing for what happens after that, which is still down to the content.
• Bigger gains on useful videos. How-to, review and explainer content responds more than entertainment.
• Most of the outcome is elsewhere. The models explained roughly a sixth of why some videos do better, so the thumbnail is a real lever and a small one.
• Default frames are costing you. The automatic frame a platform grabs follows the same rules and usually lands badly on both.

Where it does not work

• One platform, one country. All the real-world data came from Bilibili in China, where interfaces and tastes run busier than in Europe or the United States.
• Machines judged the pictures. Detail and harmony were scored by trained models rather than by people, so the measures are consistent but not quite what a human eye would call busy or tidy.
• A month of data. Videos were tracked for 31 consecutive days, which says nothing about how a thumbnail performs over a longer life.
• Made-up thumbnails in the experiments. The causal tests used nine constructed images on a simulated page, not real videos competing in a real feed.

Questions marketers ask

Also new this week

What the researchers did

The cheapest performer in advertising right now is an invented one, filmed talking about a product nobody ever bought, in the style of a customer who cannot wait to tell you about it. Jing Yang and Susanna Lee, both at Boston University, went looking for what happens at the moment the illusion fails.

They built two of these ads on a commercial platform, using the same AI avatar, a woman in her twenties or thirties. One was a 15 second vertical video about customisable family portraits, loose enough to pass as lifestyle content. The other was a 30 second piece for a real deodorant brand, built like a straight product review. Both were deliberately good rather than flawless, which a separate check confirmed.

Then they sorted viewers by how they found out. Some never twigged. Some worked it out on their own, which the authors call natural recognition. Some were told afterwards. A third study asked whether a short briefing beforehand changes how many people spot it, which is the closest thing here to a forecast.

People

187, then 196, then 238 American adults, each study with fresh participants

Ads

Two vertical videos made with the same AI avatar, one for family portraits and one for a deodorant brand

The three routes

Never noticed, worked it out unprompted, or told after watching

Measured

How surprised they were and in which direction, whether they felt betrayed, what they thought of the ad, and whether they would buy

Third study

Half the viewers were briefed beforehand that brands make ads this way

More than half worked it out

Start with how leaky this format already is. A clear majority of the first study's viewers worked out that the presenter was synthetic with no prompt at all. Not a minority of experts, ordinary people watching a short video.

What happened to those people is the finding. Their opinion of the ad sat at 2.83 out of 7, while viewers who took the presenter at face value gave 4.59, and intention to buy dropped alongside it. The gap in attitude was the larger of the two, which fits the authors' account of the mechanism: the ad broke a promise rather than failing to sell.

That promise is specific to the format. A polished brand film is expected to be staged, so nothing is violated when it looks it. A person talking into a phone about a product they bought is supposed to be a person, and viewers who discover otherwise report feeling lied to, not merely unimpressed.

Being told is softer than finding out

The second study is the counterintuitive one. Viewers who missed it entirely and were then informed registered a bigger jolt, rating the revelation as further from what they expected, 4.08 against 3.57. By the logic of surprise alone they should have reacted worse. They did not.

The reason sits in a second measure. People who were told found the whole thing considerably more novel, and novelty pulled the sting out of the surprise. Told about it, the reaction is closer to "I had no idea they could do that". Catching it yourself produces something nearer to "they thought I would not notice".

The practical reading is uncomfortable for anyone planning a campaign around this. The gentler path depends on the audience finding the technique impressive, which is a wasting asset. Novelty by definition does not survive becoming normal.

Every month makes it worse

The third study is the one to show your head of brand. Half the viewers read a short description of how advertisers make these videos, plus a post from someone who had made one, before they watched anything. That briefing alone took the share who spotted the avatar from 46% to 58%.

Nothing about the ad changed. Only what the audience knew. The authors treat this as a stand in for the real world, where the same knowledge arrives for free through news coverage, platform labels and people posting about it.

So the economics of the tactic point one way. The pool of viewers who will take your synthetic customer at face value is the largest it will ever be today, and the penalty you pay when the rest of them catch on is already measurable.

Where this hurts most

The damage is concentrated in formats that promise a real human experience. Testimonials, reviews, unboxing, influencer endorsements, before and after stories. Those are precisely the places where AI avatars are cheapest to use and most tempting, and they are the places where being caught costs the most.

Stretching the finding past what was tested, the same logic should hold wherever authenticity is the product rather than the packaging. A founder's note that was never written by the founder, a customer service reply signed by a name with nobody behind it, a case study built from a composite client. None of these were studied here, and all of them carry the same structure: a promise of a person, and a reveal waiting to happen.

The authors' own recommendation is blunt and worth repeating. If you are going to use synthetic people in a format that implies a real one, say so before the audience finds out, because the version where they tell you is the expensive one.

What to expect

• A majority will notice sooner than you think. Most of the viewers in these tests identified the avatar unprompted, with no special expertise.
• Attitude falls further than intent. The hit to how people feel about the ad is larger than the hit to whether they would buy, so brand tracking will show it before sales do.
• Disclosure buys you a milder reaction, not a free pass. Telling people still surprises them. It just lands as interesting rather than deceitful.
• The cushion is temporary. The softer response depends on the technique still feeling new, which it will not for much longer.

Where it does not work

• American panel participants only. All three studies recruited United States adults online, and detection rates vary with familiarity, which varies by market.
• Who noticed was not assigned. People sorted themselves into spotting it or not, so those groups may differ in media literacy or scepticism as well as in what they saw.
• Two ads, one avatar. Both videos used the same synthetic presenter, and recognition rates depend heavily on the quality of the specific render.
• One viewing, in a survey. Nothing here tells you what repeated exposure does, or how a brand recovers after being caught.

Questions marketers ask

Also new this week

What the researchers did

A personal note before the research. My birthday fell on the Sunday just gone, and birthdays make you sentimental about where you started. Most of my own work has circled one question: what happens to a brand when the people talking about it are not the brand. So this week's paper landed well.

Jonathan Beck (University of Mississippi), Clay Voorhees (University of Alabama), Paul Fombelle (Northeastern University) and Katherine Lemon (Boston College) have put a name to something that has crept into practice without much scrutiny: firm-generated user content. Not an ad, and not a customer's own words either. It is the post the brand writes and the customer sends from their own account.

They ran six studies. Two were live campaigns that tracked whether people actually clicked through to post. One measured what a thousand people really published on X after being handed a suggested line. The rest worked out why, by varying how enthusiastic the supplied wording was and how happy the customer had been.

Field tests

A gym emailing 966 active members, and 999 people shown a film trailer

Experiments

Scenarios covering a wine aerator, a food delivery order and a suitcase, with 1,300 participants in total

Live posting study

1,003 people watched a deliberately poor short film, then had their real posts tracked

What was varied

Blank box or ready-made text, how warm that text was, and whether the experience had gone well

What was measured

Whether they shared, what they edited, and the tone of what finally went up

A blank box is the obstacle

The first test was a real campaign. A city gym emailed its members asking them to say something supportive on social media. Half got a link to an empty posting box. Half got the same email with a finished sentence recommending the gym, ready to send. Click-throughs went from 0.4% to 2.1%.

Those numbers look tiny until you remember what they are: click rates on a marketing email, where anything above 3% is unusual. Five times as many people took the step. The second field test, built around a film trailer, moved the same way, from 1.4% to 3.4%.

Nothing here is about persuasion. The customers who shared already liked the gym. What changed was the cost of doing something about it, and most goodwill never converts because writing a sentence in public is more effort than it looks.

Enthusiasm is the dial

So what should the supplied sentence say? The second study put four versions in front of people who had just had a good experience: an empty box, a flat statement of fact, a warm line, and an outright gushing one.

The flat version is roughly what brands write today, and it barely beat the blank box. Both of the warmer versions nearly doubled sharing. The difference between warm and gushing was small, which matters, because the gushing one carries a cost that shows up in the next section.

A second effect runs alongside. People overwhelmingly send the supplied text as written, so whatever tone the brand chooses is the tone that ends up in public. Where customers wrote their own words, the posts came out noticeably cooler.

The unhappy customer quietly declines

Now the same four versions, sent to people whose experience had gone badly. The pattern inverts. The glowing line, which worked best on happy customers, cut sharing to roughly a third of what a blank box produced.

The reason is simple enough to recognise from your own inbox. For a satisfied customer, a ready-made line is a favour: it says what they already think. For a dissatisfied one, it is a sentence that does not fit, and the mismatch is uncomfortable enough that most people drop it rather than rewrite it.

The live study is the reassuring part. Of the unhappy viewers who did go ahead and post, four in five sent the supplied line untouched, and the fifth softened it rather than sharpening it. Nobody turned it into an attack. So the fear that this hands angry customers a megaphone did not materialise; what it does instead is thin out their presence in the feed.

The part that should make you pause

Here is where I have to take my marketer's hat off. The entire value of user-generated content rests on readers believing it came from a user. It persuades precisely because it is not controlled by the brand, which is why my own work has always treated firm-created and user-generated communication as two different animals that do different things to brand equity.

A tactic that works this well erases that line. Scale it across a customer base and the brand's mentions stop being evidence and start being distribution. The authors raise the same worry, and note that it may eventually call for the kind of labelling that sponsored posts already carry.

Going beyond what their studies tested: the practical limit is probably self-correcting. If every mention of your brand reads like the same sentence, readers will notice, and the channel will lose the credibility you were borrowing. Which argues for using this where the enthusiasm is genuine, varying the words, and treating the number of supplied posts against spontaneous ones as a metric worth watching rather than maximising.

What to expect

• Small percentages, real volume. Share rates move by a point or two, which against a large customer base is a meaningful lift in mentions.
• A warmer public record. Supplied text travels mostly unedited, so the average tone of what gets posted rises.
• Fewer negative posts, not more. Unhappy customers tend to opt out rather than rewrite, and those who do post soften rather than attack.
• A credibility budget you are spending. Every supplied post trades a little of the independence that makes this channel work.

Where it does not work

• Short posts only. This was tested on single-sentence social posts, not reviews, ratings or anything long enough to need a personal voice.
• One suggestion at a time. Every test offered a single ready-made line; offering a menu to choose from was not tried.
• No long-run evidence. Nothing here shows what repeated asks do to trust in the brand or in the channel over months.
• Mostly US scenario-based samples. Four of the six studies asked people to imagine an experience rather than live one, with the two field tests run in a single country.

Questions marketers ask

Also new this week

What the researchers did

YouTube, X and TikTok have all tested letting people register a thumbs down, and all have worried about the same side effects. Varad Deolankar (National University of Singapore) with Jessica Fong and S. Sriram (both University of Michigan) went looking for evidence on a platform that has had the button for years.

They picked 50 subreddits, took the 500 most recent people to start a thread in each, then tracked everything those users did anywhere on Reddit for 61 days over the turn of 2024. That gave 1.9 million comments spread across about 32,000 communities, each one's score checked daily for a fortnight after it was posted.

The hard part is proving cause. People who are fired up post more and attract more downvotes, so the two move together for reasons that have nothing to do with the button. The authors found a way round it borrowed from pricing research: the same loss feels bigger when the leading digit changes. A reputation score sliding from 103 to 99 stings in a way that 104 to 100 does not, even though both lost four points. Comparing those two cases isolates the effect of noticing.

Platform

Reddit, 23 December 2023 to 22 February 2024

Users tracked

17,525, followed across every community they posted in, not just the ones they were found in

Content

1.9 million comments; 97.6% of everything people posted was a comment rather than a new thread

How often this happens

Of the comments that got any reaction on a given day, 48% were being pushed down

The trick

A change in the leading digit of a user's score, which makes the same loss feel larger

Nobody went quiet

The average person in the sample commented on 29% of days. After a knock to their score that crossed one of those digit thresholds, that figure rose by 8.5 points, and the number of comments they wrote rose too. Praise worked almost identically: a noticeable gain added 9.3 points.

That symmetry is the interesting bit. Being pushed down and being pushed up are not opposites here. Both are evidence that somebody is out there reading, and both pull people back to the keyboard the following day.

A placebo check supports the reading. When a score drifted within five points of a threshold without crossing it, nothing happened. The reaction is to the loss being noticed, not to its size.

They move along, not out

The fear behind the dislike button is that it sorts people into rooms where everyone already agrees. If a downvote drove people away from the room where it happened, the button would be an echo-chamber machine.

The data splits the extra posting three ways: back into the same thread, into the same community but a different thread, or into a different community altogether. Only the third moved. Crucially, the first two did not drop either. People did not retreat from the room that had just told them they were wrong; they just also went and said something somewhere else.

For anyone running a brand community, a forum or a review section, that is the reassuring half of the paper. A negative signal is not an exit signal.

The loudest posts came back quieter

Volume is one thing, tone another. The authors ran every comment through Google's language API, which picks out the topics in a piece of text and scores how strongly the writer feels about each one. Then they followed each user to the next time they wrote about the same topic.

Mild posts that got downvoted came back slightly sharper, as if the writer had decided to make the point properly this time. Strong posts came back softer, and the stronger the original, the bigger the climbdown. The turning point sits at a level of intensity that about a third of all comments exceed, so this is not a corner case.

It also held in the places where you would least expect it. The authors sorted some 98 subreddit topics into controversial and ordinary, using an AI classifier, and the softening was at least as strong in the political and contentious ones.

What the button is really doing

One test gives the game away. When a user's score dipped below a threshold and then climbed back above it the next day, the extra posting disappeared. People were not learning what the crowd wanted; they were repairing a number they could see.

That reframes the finding for anyone who runs a scoring system of their own, which is most of us now: seller ratings, driver scores, loyalty tiers, credit-style dashboards, streak counters, internal league tables. Push the argument one step past the evidence and the lesson is less about negativity than about visibility. A score has to be public and cumulative before a dent in it is worth repairing, and where the number sits relative to a round figure changes how big the dent feels.

The practical corollary is uncomfortable. If you want the behaviour, you have to let people feel the loss. If you cushion every setback, smooth the score, or hide it behind a weekly summary, the thing that drove the response is gone.

What to expect

• A next-day effect, not a lasting one. The lift showed up the day after and the day after that, then faded.
• Praise and criticism pull in the same direction. Both increased activity, so a dislike button is not the mirror image of a like button.
• The tone effect depends on where you start. Mild contributors sharpened up slightly; strong ones backed off.
• Most reactions are ordinary. Only a small share of scores ever cross a threshold or flip sign, so the dramatic cases are rare even when the average effect is real.

Where it does not work

• One platform, one window. Reddit over 61 days, among users picked for being recently active, which skews towards the committed.
• Only the net number is visible. Reddit shows upvotes minus downvotes, so nobody, the authors included, can tell ten downvotes from eleven upvotes and twenty-one downvotes.
• Votes only, no words. Replies and private messages were not observed, and they may be doing some of the work attributed to votes here.
• Polarization was not measured. The study tracks how one person's tone changes, not whether the conversation around them calms down.

Questions marketers ask

Also new this week

What the researchers did

A deepfake puts a face into footage it was never in. Malgorzata Karpinska-Krakowiak (University of Lodz) and Martin Eisend (European University Viadrina and Copenhagen Business School) set out to test whether the labels the industry reaches for protect anybody at all.

They built their own deepfaked ads. In the first, a perfume commercial starring Charlize Theron was altered so that Margot Robbie appeared to be the one on the catwalk, selling an invented brand. In the second, a food delivery ad fronted by Kim Kardashian had her face replaced with a stranger's. Then they attached different wording under each ad and put it in front of fresh samples of British adults.

The clever part is what they measured. Earlier research on retouched photographs asked whether a label makes an image look less realistic, and the answer is almost always no. The authors split that idea in two: how real a thing looks, and whether people believe it actually happened. A deepfake is designed to be the first without being the second, so the second is the only one a label can hope to shift.

Studies

Three online experiments, with fresh samples each time

People

121, then 308, then 300 UK adults recruited through Prolific

Ads

A deepfaked perfume ad, a deepfaked food delivery ad and a deepfaked chocolate ad, all for invented brands

Labels tested

None, a plain statement of the swap, that statement plus what it means for the ad, and that plus what it means for the viewer

Outcomes

How realistic the ad looked, whether it was believed to be truthful, liking of the ad, and whether people clicked to buy

The fake is invisible until you say so

First the uncomfortable baseline. When the deepfaked perfume ad carried no label, viewers rated how synthetic it looked at 3.61. The untouched original scored 3.73. In other words, people were no more suspicious of the doctored ad than of the real one, and if anything slightly less. Only when a disclaimer appeared did the rating jump to 5.71.

So detection is not happening on its own. Whatever protection exists has to come from the words underneath.

Then the bad news about those words. The label used in this first study was the sort you see in the wild: a byline explaining that the post contains synthetically generated images, with some elements swapped in from other sources. It made no difference to anything that mattered. Realism came out at 4.85 with the label and 4.84 without. Liking of the ad, and the decision to click through and buy, were equally untouched.

Finish the sentence

The second experiment is where the paper earns its keep. Four versions of the same deepfaked delivery ad went out. One had nothing underneath. One had the plain statement: the face of the original actress has been swapped with somebody else's. A third added a consequence for the ad itself — so it does not portray reality, it is fake. A fourth added a consequence for the viewer — the goal is to mislead you into thinking you are watching reality.

Belief in the ad's truthfulness barely moved between no label and the plain one, 3.33 against 3.20. Both of the versions that carried a consequence pulled it down sharply, to 2.41 and 2.68, and they were not meaningfully different from each other. The third study repeated the pattern with a different product.

How realistic the ad looked never shifted in any condition, exactly as predicted. A well-made fake looks well made whatever you write beneath it. What you can change is whether people think it happened.

It does not cost you the sale

Here is the objection every marketer raises, and the answer the authors anticipated. If a disclosure teaches people the ad is untrue, surely it wrecks the ad.

It did not. In the second study, the share of people who clicked to buy ran from 13.5% with no label to 17.9% with the plain one, 15.8% and 15.0% with the two stronger ones. Statistically that is a flat line. The third study, with a chocolate brand, produced the same flat line at a higher level.

Underneath, two forces were cancelling out: believing the ad less pushed liking down and buying with it, while the disclosure itself nudged choice up a little, perhaps because candour reads as confidence. The authors treat that cancellation as the point rather than a disappointment. A label that informs people without hurting the campaign is one a brand will actually agree to run.

Who the label leaves behind

The third study added the catch. The stronger labels worked best on people who expected advertising to tell the truth in the first place. On people who assume ads are embellished anyway, the warning landed on ground that was already conceded and changed little.

That is an awkward result for policy. The consumers best served by transparency rules are the trusting ones; the cynical are left where they were, which is not the same as being protected. Nor does scepticism help you spot a deepfake, as the first study showed.

Reaching past what these experiments tested, the lesson generalises to every label now being bolted onto synthetic media: the AI tags on social platforms, the small print under a cloned voice in a radio spot, the badge on a virtual influencer's profile. Naming the technique tells people what was done. It does not tell them what to do with that knowledge. The sentence that changes behaviour is the one after the disclosure, and almost nobody writes it.

What to expect

• No change in how real it looks. Perceived realism held steady across every label in every study, so do not judge a disclosure by whether the ad still looks convincing.
• A shift in belief, not in behaviour. The wording moved what people thought was true; clicking to buy stayed flat.
• Diminishing returns on words. The consequence for the ad did the work; adding the personal warning on top was no better.
• Smaller effects on sceptics. Audiences who already assume advertising is embellished respond least to being told it is.

Where it does not work

• Invented brands, single exposures. Participants saw one still social post for a fictitious product, not a campaign from a brand they have feelings about.
• British samples only, and one of them all women. The first study ran with 121 women aged 18 to 60; the later two were mixed but still UK-based Prolific panels.
• Belief in these ads was low to start with. Perceived truthfulness sat below the midpoint everywhere, so the labels were pushing an already sceptical audience further.
• Clicks in a survey, not purchases. Choice was a two-button decision inside the questionnaire, with no money involved and no time to reconsider.

Questions marketers ask

Also new this week

What the researchers did

Disclosure before we start: I am one of the four authors, so take the enthusiasm with a pinch of salt.

Davide Marengo (University of Turin), Halley Pontes (Birkbeck, University of London), Christian Montag (University of Macau) and I wanted to know whether the kind of game someone loves says anything about how much trouble their gaming causes them. The survey ran in 2019 through the Electronic Sports League as part of a healthy-gaming campaign, with no payment on offer; players got a personal read-out of their own habits instead. After cleaning, 116,047 people were left.

Everyone named their current favourite game in a free-text box. Those titles were matched to RAWG, a public database of more than 618,000 games, which supplied the genre tags and a yes/no flag for multiplayer. Each player also reported the device they mostly use, how long they play, and answered two standard questionnaires on disordered gaming, one built on the American psychiatric criteria and one on the World Health Organization's.

People

116,047 players, from an opening pool of 188,775 responses across 181 countries

Who they were

Average age 24.1 (range 12 to 80), 92.9% male, 51.6% educated beyond school

Time on the hobby

23.3 hours a week on average

Devices

Desktop 57.0%, console 27.1%, laptop 12.3%, phone or tablet 3.5%

Game classification

Genre tags and a multiplayer flag pulled from the RAWG database

Four screens, four shopping lists

Start with the common ground. Action tags sat on most of the favourites named on every screen. Below that line the four groups pull apart.

Console players leaned into things you can watch someone do: shooters (36%), adventure (31%), sports (9%) and fighting games (6%), the last at more than four times the rate found on any computer. Desktop and laptop players went for worlds that keep running: massively multiplayer games (30% and 25%), role-playing games (43% and 46%), indie titles (12% on both) and strategy, which climbed from 4% on consoles to 26% on desktops and 35% on laptops. Phone and tablet players were the quiet opposite: arcade (11%), casual (11%), card games (5%) and puzzles (4%), each several times higher than anywhere else.

These are the favourite titles of real people, not a panel's claimed interests, which makes the contrast useful. A sponsorship, a product placement or an in-game media buy that treats all four groups as "gamers" is buying four different rooms with one brief.

Age lives in the genre, not in the channel

Marketers often reach for gaming when the brief says young. The data inside the channel is more interesting than that. Older players in this sample drifted towards board games, card games, puzzles, adventure and role-playing titles; younger ones towards action, arcade and shooters, and towards anything with other people in it.

Education tracked the same line. Players with a degree were more likely to pick board, card and puzzle games and less likely to pick action, arcade and shooters. Women, who made up 7% of this particular sample, showed a mild pull towards casual games and nothing else worth reporting.

So the age read comes from the genre, not from the fact that someone games at all. A puzzle app and a shooter sit in the same industry and sell to different decades.

Almost nobody is playing alone

Multiplayer was the rule rather than the exception: 78% of the favourite games had it. The habit was strongest on computers and weakest on phones, and it was not a fringe of hardcore titles either, since even casual and card games now routinely ship with someone else in them.

It is also where the weight sits. Players whose favourite game had multiplayer scored higher on both disordered-gaming scales, and on consoles the pattern was sharpest, with multiplayer, massively multiplayer, casual and shooter preferences all tilting upwards. Puzzles and platformers tilted the other way. The differences were small in statistical terms, but they ran in the same direction under both diagnostic systems and across 116,047 people.

For anyone buying attention in games, that is a double message. The audience you most want is sitting in social, never-ending titles, and those are the same titles where a minority of players are struggling. Pushing spend or urgency at that group is a risk worth naming out loud in the plan.

Past the console

Here is the finding that travels furthest. When the team fed every genre tag and the multiplayer flag into a machine-learning model and asked it to predict an individual's disordered-gaming score, it managed almost nothing: a correlation between .08 and .15 between predicted and actual scores. The labels describe the crowd well and the person badly.

Reading past what the paper set out to test, that gap is the shape of most audience data. Device, genre, platform, channel: these are strong group facts and weak personal ones. They are good enough to decide where a campaign runs and what it looks like, and nowhere near good enough to decide what one customer is like, how much they can afford or whether they are in trouble.

The same split applies well outside games. Streaming services, fitness apps, trading platforms and betting products all sort their users by the surface they use and then draw conclusions about individuals from it. The first half of that move is sound. The second half, on this evidence, is wishful.

What to expect

• Large gaps between devices, small gaps between people. Genre splits of 20 or 30 points are normal here; the differences in individual behaviour behind them are slight.
• Mobile is a different business. Arcade, casual and card games were several times more popular on small screens than anywhere else, and shooters far less.
• Your own mix will not match this one. The sample is 93% male and PC-heavy, recruited through competitive gaming.
• Both measuring systems agree. The WHO and American criteria produced the same pattern, so the findings do not hinge on which questionnaire you trust.

Where it does not work

• A 2019 snapshot. Data were collected before the pandemic; mobile gaming in particular has changed a great deal since.
• An esports crowd, not the market. Recruitment through the Electronic Sports League drew competitive PC players and very few casual ones.
• Blunt tags. Genre labels and a yes/no multiplayer flag come from a community database, so a game with an optional co-op mode counts the same as a persistent online world.
• A snapshot, not a cause. Everything is self-reported and measured at one moment, so nothing here shows that a genre creates heavy play rather than attracting it.

Questions marketers ask

Also new this week

What the researchers did

Cashierless shops, robot hotel desks and smart-shelf supermarkets are spreading fast, and most studies of them ask one thing: do customers like them? Usually they do. Xiaoyan (Jenny) Liu (Southwestern University of Finance and Economics), Chi Hoang (ESCP Business School) and Sharon Ng (Nanyang Technological University) asked a different question: what happens to people after they walk out?

Their focus was everyday kindness: writing a note to an orphan, volunteering a few hours, dropping money in a charity tin. They ran nine experiments, three of them in real places, comparing an AI-run version of a store, hotel or restaurant with the same place run by staff. Then they tested what brings generosity back.

Studies

Nine experiments, including field studies at two neighbouring convenience stores and a campus hotel in Singapore

People

About 3,200 participants across four countries

Settings

Grocery and convenience stores, a hotel with a humanoid robot, and a robot-staffed restaurant

Generosity measured by

Letters written, volunteer hours pledged, donation amounts and choosing a longer, more helpful survey

Fewer people, less giving

The neatest test happened on a university campus in Singapore, where one retail brand runs two convenience stores side by side, with the same layout and the same prices. One has staff; the other has none. Researchers waited outside both and invited shoppers to sign up as student buddies. People coming out of the staffed shop offered far more of their time, a gap that held after accounting for age, gender and shopping habits.

Online, the pattern repeated. Americans who imagined a trip to an AI-equipped grocery store wrote shorter encouraging letters to children supported by a charity than those who imagined a staffed store or simply recalled their last shop. A follow-up hotel study compared self-service check-in kiosks, a robot and a human desk: both machine versions led to smaller donations, and the robot was no improvement on the kiosk.

Why: the missing small talk

A quick hello at the till seems trivial, but it reminds us we're part of a community. In a UK study, people who pictured an autonomous grocery store felt far less connected to others (2.83 against 4.64, out of 7), and were less keen to buy a £1 felt flower made by orphans for charity. That loss of connection, not a sense of threat from the technology, explained the drop.

Two checks back this up. When participants were first nudged to think about how much they rely on other people, a robot-run restaurant no longer dented their wish to volunteer for the Salvation Army. And people didn't simply dislike the automated stores: they often rated them as highly as, or higher than, the staffed ones.

Three ways to win the warmth back

Shop together. Chinese participants who pictured shopping alone in an AI store were less willing to volunteer with older neighbours (7.08 against 7.69, out of 9). Picture the same trip with family, and the gap disappeared.

Say who the technology is for. Back at the Singapore stores, some shoppers in the autonomous shop saw a short message, adapted from the retailer's own press release, saying the system was introduced to make shopping faster and easier for customers. Their volunteer pledges climbed back to the level of the staffed store. A UK replication with donations in pounds showed the same thing.

Ask donors what's in it for them. A charity appeal stressing benefits to the whole community lost ground after an AI store visit (6.79 against 7.39 out of 9). The same first-aid-kit project, pitched as protecting "you and your family", did not.

Far from the checkout

The authors studied shops, a hotel and a restaurant, but the same trade-off shows up anywhere a person has been replaced by a screen. Taking our own step past the data: app-only banks, contactless hotel stays, parcel lockers, chatbot-only customer service and self-checkout lanes all remove the small human moments this study points to.

That matters for more than charity tins. Brands that run cause campaigns, round-up-at-the-till schemes or community programmes through automated channels may be asking for generosity at exactly the moment people feel least connected. And for retailers who sell themselves as "part of the neighbourhood", the human touch may be doing more work than the efficiency case suggests.

What to expect

• Small dips, many customers. The drop per person was modest, but across thousands of daily visits it can noticeably shrink donations and sign-ups.
• Satisfaction surveys will miss it. Customers rated the automated stores just as well, so the cost hides from the usual dashboards.
• Charity partners may feel it first. The authors expect collection tins and sign-up desks next to autonomous stores to do worse unless the appeal is reworked.
• Fixes are cheap. A message, an appeal tweak or a group offer costs little compared with restaffing a store.

Where it does not work

• Mostly imagined shopping trips. Six of the nine studies asked people to picture a store; only the Singapore studies used real ones.
• Young, local field samples. The real-store studies involved university students on one campus.
• Short-term only. Generosity was measured minutes after the visit; whether the effect lasts, or fades with habit, is unknown.
• Mostly intentions. Several measures were stated willingness to give rather than money actually handed over.

Questions marketers ask

Also new this week

What the researchers did

Ask people whether they'd rather be served by a person or a machine and most pick the person. Yeseul Kim (University of Bath), Dipayan Biswas (University of South Florida), Courtney Szocs (Penn State) and Khalia Jenkins (American University) wanted to know who the exceptions are, and when.

Their starting point is uncomfortable but well documented: in US banks, shops and rental offices, Black customers often wait longer, get less warmth and face more suspicion. Being the only non-White person in the room can bring all of that to mind. The question was whether, in that moment, a machine starts to look safer than a person.

The team ran ten studies. Two were real job adverts with paid work at the end, one was a Facebook ad campaign, and the rest were online experiments set in banks, rental offices and an investment firm. Every time, people chose between a human and an algorithm: an AI interviewer, a robot teller or a digital kiosk.

Studies

Ten US studies with nearly 4,000 participants, plus a Facebook ad campaign

Settings

Banking, apartment leasing, investment advice and job interviews

Choice

A human employee or an AI interviewer, robot teller or digital kiosk

Surveys

97 non-White consumers and 201 managers on concerns about being judged in service settings

Identity, not technology, drives the switch

The first job advert, posted at historically Black colleges, set the pattern. A second run, shared through Black community networks, showed an even steeper jump: from 15% to 37% choosing the AI interviewer. Online, Black participants looking at a bank full of White customers chose the robot teller 38% of the time; in a mixed crowd, 22%.

What tipped the balance was a specific worry. When non-White participants were told they felt out of place at a bank, their concern that staff would size them up by race rose from 2.57 to 3.79 on a 7-point scale, and that concern explained the move towards the robot.

It isn't just people wanting staff who look like them. White men in the bank study said they felt less similar to non-White employees, yet they went to the kiosk no more often (38.8% against 37.5%). Feeling different only mattered for people who had reason to expect unfair treatment.

The same pattern in real ads

The Facebook ads ran for one day at a time in mostly Black and mostly White US neighbourhoods. Where most residents were Black, a diverse advice team did almost as well as the chatbot. Where most were White, the ranking flipped: the all-White team drew the most clicks and the chatbot fell behind. One image, two very different audiences.

Signals of inclusion change the picture

The pull towards machines weakened wherever judgment felt less likely. Applying to lease a flat, people who had just recalled a time they stood out as a minority chose an AI interviewer 52% of the time, against 36% in the control group. Applying to a social club for people with mixed interests, that same reminder made no real difference.

Telling people the bank was minority-owned did something similar. The jump towards the robot teller shrank to eight points (24% to 32%), against twenty-two (30% to 52%) without that line. And at the investment firm, a short diversity statement was enough to cancel out the effect of an all-White team.

Beyond the bank branch

The authors studied race and ethnicity in the US, and they suggest that other groups who expect to be judged, by sexual orientation or by wealth for example, may react the same way. Going a step further than the paper, the same logic could hold at a GP's front desk, an insurance claims line, a debt advice service, a car dealership or a luxury boutique: anywhere a customer might feel assessed as a type rather than helped as a person.

It also changes how to read your channel data. If one group of customers sticks to the app, the kiosk or the chatbot, convenience may not be the whole story. They may be telling you something about the counter.

What to expect

• Most people still want a person. Even in the strongest conditions, about half of the customers stayed with the human option.
• Shifts of roughly 10 to 20 points. That was the usual size of the swing towards the machine when customers felt they stood out.
• The payoff is in high-stakes services. Banking, renting and hiring showed the effect; a relaxed social setting did not.
• Cheap to try. An extra channel and a visible statement cost far less than restructuring a team, and both are easy to A/B test.

Where it does not work

• US, race and ethnicity only. Other identities and countries were not tested.
• Mostly hypothetical choices. Only the job adverts and the Facebook campaign recorded what people actually did.
• Perceived fairness, not actual fairness. The study didn't check whether the AI options treated people more equally.
• Some results were marginal. A few comparisons only just reached significance, and click-through rates on the ads were very low.

Questions marketers ask

Also new this week

What the researchers did

Most marketers treat personalization as a safe bet: it saves customers effort and makes offers more relevant. But every personalized message is also a mirror. It tells customers what the brand knows about them, and sometimes that is something they would rather not see. Seven in ten Spotify users, the authors note, say they are embarrassed by their annual "Wrapped" summary.

John Yi (Le Moyne College) and Caleb Warren (University of Arizona) tested when that mirror backfires, using three studies with different kinds of personalization: an online pharmacy, a music streaming service and a real Google display ad campaign. In each, they varied whether the marketing was presented as personalized and whether it pointed to something people would rather not be associated with.

Study 1

398 US adults with a BMI over 30 shopping at a mock online pharmacy for weight-loss or headache medicine

Study 2

291 US undergraduates using a test music service that recommended playlists

Study 3

A 12-day Google display ad campaign shown to people who had searched for acne terms: 324,014 impressions

Measured

Ratings of the business (1 to 7), embarrassment, privacy concern, and clicks in the field

The online pharmacy

Participants returned to an online pharmacy to buy either weight-loss medicine or headache medicine. Half were greeted with a personalized page: "Based on your last purchase, we think you will need this. It's time for your monthly refill! Personalized recommendations, just for you." The other half found the product by browsing the menu.

For headache medicine, the greeting made no difference. For weight-loss medicine, it backfired: people rated the pharmacy 3.02 out of 7 with personalization against 3.79 without, even though it saved them clicks. The reason was embarrassment. Concern about privacy and about being identified did not explain the drop.

The playlist and the ad

In the second study, students chose between artists, then saw two recommended playlists with exactly the same songs. Only the names changed: "Almost Sweet 16" and "Tween Party Music", or "Almost 21" and "College Party". When the service said the playlists were personalized, students liked it less if the names suggested tween taste (4.47 against 5.13 out of 7). With the college names, personalization did no harm.

The field test ran Google display ads to people who had searched for acne terms. One ad offered "10 ways to fix your acne"; the other, "10 ways to maintain your perfect skin". Each said it was either "shown to you based on your past Google searches" or shown to a random selection of visitors. Saying the acne ad was personalized cut clicks by 28.5%, while saying the clear-skin ad was personalized raised them by 31%. The difference between these two patterns was statistically reliable, though each comparison on its own was not, as clicks were very rare.

Beyond the pharmacy shelf

The study tested medicine, music and skin care, but the risk applies to any category people do not want reflected back at them. My reading, not tested in the paper, is that this includes debt and payday loans, hair loss, hearing aids and incontinence, alcohol, gambling, dating, diet products and guilty-pleasure viewing, and also year-in-review recaps, "because you watched" rows and named email subject lines that can be seen on a lock screen.

The flip side matters too. Personalization that points to an identity people are proud of, healthy, stylish or green, can work in your favour, as the clear-skin ad suggests. The question is not whether to personalize, but when to say so.

What to expect

• No penalty in neutral categories. Personalized service for headache medicine, and "College Party" playlists, did no harm.
• A clear drop in sensitive ones. Ratings of the pharmacy fell by about three-quarters of a point on a 7-point scale for weight-loss medicine.
• A possible boost for flattering identities. The personalized clear-skin ad drew 31% more clicks, though that result was only marginally reliable.
• Privacy fixes will not solve it. In Study 1 the drop came from embarrassment, not privacy concern, so better data policies alone will not help.

Where it does not work

• Mostly imagined or lab settings. Two of the three studies used simulated shopping or a test service, with US adults and students.
• Little real benefit from personalization. The personalized versions saved at most a click or two, so the studies do not show what happens when personalization offers real value.
• A weak field signal. The ads drew only 249 clicks from 324,014 impressions, and the individual comparisons were not statistically reliable.
• Some effects were marginal. In Study 1 the overall interaction just missed the usual significance threshold, though the weight-loss comparison itself was clear.

Questions marketers ask

Also new this week

What the researchers did

Contextual advertising, placing an ad next to content that suits it, is a market worth more than 200 billion US dollars a year. In practice, ads are usually matched to content by product and media category: fashion ads in lifestyle titles, kitchen ads on cooking pages. Jianping Ye and Michel Wedel (University of Maryland) and Rik Pieters (Tilburg University and Católica Lisbon) asked whether that is the best way to win attention.

They used one of the largest eye-tracking databases studied so far: 3,531 display ads from 29 Dutch digital magazines, each viewed by 80 to 100 people as they leafed through the pages. From images of each ad and the page opposite it, they extracted dozens of features, from size, colour texture and visual clutter to the topics shown, described by an AI model that reads both words and pictures. A machine-learning model then learned to predict how long people looked at the ad and at its brand. Finally, they tested which placements would win the most attention.

Data

3,531 display ads in 29 digital magazines, 71 product categories and 812 brands; about 3,600 Dutch consumers

Measured

Seconds of eye gaze on the whole ad and on its brand name or logo

Checks

55 very different ads (banners, native, outdoor) and a controlled experiment with 715 people

Tool

AdGazer, with free code and a public web app that predicts gaze from an ad and page image

The page around your ad matters

The model predicted attention well: its predictions correlated .83 with real gaze on ads and .80 with gaze on brands, better than two widely used image-recognition models. It also held up on ads it had never seen, such as banners, native ads and outdoor posters.

The size of the ad and of its brand, text and picture still mattered most. But features of the surrounding page accounted for 33% to 43% of predicted attention to the ad, and about 20% of attention to the brand. Counting how ad and page features work together, the context was behind about half of ad attention.

And the page competes. In a controlled experiment where 24 ads were placed at random beside editorial pages, the same ads drew about 1.2 seconds less gaze, and their brands about a quarter of a second less, than when shown alone. Good matching is a way to reduce that loss.

Topic matching alone does not win attention

The researchers then reassigned 749 ads to the available pages in 95 magazine issues, using the model to predict attention for each pairing. The placements actually used in the data were no better than random. Re-placing ads by product and media category, or by text and keywords only, the way most contextual targeting works, did not improve on them either.

What worked was visual fit. Matching only on how ad and page look, the logic of native advertising, raised predicted attention by about 5% when any magazine could be chosen. Using both look and topic together raised it by 4.8% within a single magazine and by 10.1% across magazines. Choosing the magazine with the ad in mind doubled the gain.

Beyond the magazine page

The study covers digital magazines viewed on a desktop screen, but the decision it improves is everywhere: which ad goes next to which content. My reading, not tested in the paper, is that the same logic applies to retail media pages, publisher sites, in-app placements and digital out-of-home screens, wherever you can see the ad and its surroundings before you buy. Matching on what the page is about is not enough; what it looks like matters as much.

The bigger shift is cost. Predicting attention used to need an eye-tracking lab. Here each image is processed in about 1.5 seconds, so screening many ad-and-page combinations before a campaign becomes realistic.

What to expect

• Gains of 5% to 10% in attention. That is what better placement predicted here; the authors cite earlier research linking 10% more gaze to roughly 1% to 5% more sales.
• A smaller effect on the brand. Context drove about a third or more of ad attention but only about a fifth of brand attention.
• Context costs attention by default. Ads next to content got less gaze than ads alone, so a poor match is a real loss, not a neutral choice.
• Low cost to try. The code and web app are free, and scoring an ad takes seconds.

Where it does not work

• Magazines on a desktop screen. The data come from digital magazines viewed in a lab in the Netherlands; results may differ for phones, social media feeds, video or search.
• Prediction, not cause. Feature contributions show what predicts attention, not what causes it, so they suggest hypotheses to test.
• Simulated placements. The 5% to 10% gains are model predictions from re-assigning ads, not results from a live campaign.
• Attention is not sales. The study measured gaze; the links to sales and revenue come from other research.

Questions marketers ask

Also new this week

What the researchers did

Search advertising is built on two habits: people type short keywords, and then they click links. AI search engines, which write an answer instead of listing pages, could break both. Most claims about this so far have been guesses, so Gabriel Garlough-Shah (Northwestern University) and colleagues at the University of Minnesota watched what people actually do.

They built a Chrome extension that logged every search, result and click. US adults who had used both kinds of search engine in the past three months searched as they normally would for 24 hours, using Google and Google Gemini. Each search was then sorted by what the person wanted: facts (information), something to buy or a product to look up (shopping), or a recommendation or judgement (advice), such as "what should I consider when buying a car?".

People

52 US adults (average age 36), recruited through Prolific, who regularly used both kinds of search

Searches

848 real searches logged between October 2023 and November 2024, about 16 per person

Engines

Google (traditional search) and Google Gemini (AI search)

Measured

Which engine people chose, how many words they typed, and whether they clicked a result

Shopping stays on Google, advice moves to AI

Overall, 76% of searches went to Google. The split depended on what people wanted. For shopping searches, people used Google 94% of the time; for information, 76%. For advice they leaned the other way, choosing AI search 54% of the time.

Some of that advice was commercial. Questions such as "what brands are best for architectural shingles?" or "what is the best dog food for a Shetland sheepdog?" went to the AI. Product-related searches on AI search were still only about 5% of all searches, but they show the comparing-and-choosing stage of buying starting to move.

Longer questions, fewer clicks

People searched differently on AI. Queries averaged 11 words on Gemini against 4 on Google, and the gap held for every type of search: even shopping queries were twice as long on AI (7.4 words against 3.5). Short "head" terms of two words or fewer, the expensive keywords advertisers bid on most, made up 35% of Google searches but under 5% of AI searches.

Clicks dropped sharply. People clicked a result in 59% of Google searches and 12% of AI searches, so Google produced about 14 times as many clicks. On Google, shopping and advice searches led to the most clicks; on AI search, every type stayed low, because the answer itself was often enough.

What it means for search advertising

The authors draw three conclusions. First, bidding on short head terms on Google is unlikely to suffer soon, because people still type those into Google. The long tail, longer and more specific phrases, is where volume may leak to AI. Second, keyword matching fits AI search poorly: an ad bought on "best drills" matches badly with "what are the best drills for a contractor who works long days, needs to be mobile and works with dense materials". Targeting will need to follow intent, not just words. Third, with so few clicks, pay-per-click may not be the right model for ads inside AI answers.

For now, the authors say, the most practical lever is making content that AI engines pick up and describe correctly, often called generative engine optimisation, and making websites easy for AI to read.

Beyond the search box

This is not only a search-advertising problem. My reading, not tested in the paper, is that any brand content that used to earn a visit, such as buying guides, comparison pages, FAQs and how-to articles, now risks being summarised in an answer the customer never clicks through. The upside is that AI search takes the long, situation-rich questions, which reveal far more about the customer than a two-word keyword.

Google has meanwhile been putting AI answers into its ordinary results pages, so the line between "Google" and "AI search" is already blurring. The pattern to watch is the one the study shows clearly: people go to AI when they want a judgement, not a list.

What to expect

• Google remains the shopping channel for now. In this study 94% of shopping searches went to Google.
• Fewer visits from AI answers. About one AI search in eight led to a click, so visibility there will not show up as much traffic.
• Gradual pressure on the long tail. Long, specific searches are the ones moving to AI, so their volume and cost on Google may shift first.
• Less control over your message. AI engines rewrite what brands say, and the same question can get a different answer each time.

Where it does not work

• A small group. The findings come from 52 people, with heavy drop-out from the original 694 surveyed, so they show patterns, not market shares.
• One day of searching, while being watched. People knew their searches were recorded, used a fresh browser profile and were observed for only 24 hours.
• Fast-moving products. The data were collected in 2023–2024 with Gemini as a separate tool; AI answers now also appear inside ordinary search pages.
• Machine-coded intent. Each search's purpose was classified by GPT-4, which matched human coders about 87% of the time.

Questions marketers ask

Also new this week

What the researcher did

A tourist trap survives because visitors cannot tell good from bad before they sit down, and rarely come back. Reviews should fix that, but proving it is hard: restaurants that get good reviews may simply have been good already.

Dante Donati (Columbia Business School) used a natural experiment. On 15 June 2017 the EU abolished roaming charges, so European visitors could use mobile data abroad at no extra cost. Overnight, checking Tripadvisor on the street became free for them, but not for Italians or for visitors from outside the EU. After the change, EU reviewers became 1.4 times more likely to post from a phone rather than a computer.

He matched about 3 million Tripadvisor reviews to official social security and company records for restaurants across the province of Rome, so he could see revenues, staff, wages, hiring and closures, not just ratings. He then compared restaurants that tourists could easily reach with those they could not, before and after the change, separately for restaurants rated below, around and above 4 stars.

Setting

Restaurants in the province of Rome, 2015–2019, around the EU roaming abolition of June 2017

Data

About 3 million Tripadvisor reviews matched to employer and tax records; 4,628 restaurants rated before the change

Rating groups

Low (below 3.85), mid (3.85 to 4.25) and high (4.25 and above), roughly below, around and above 4 stars as displayed

Measured

Revenue, staff numbers, closures, openings, hiring of experienced workers, wages and later ratings

Good ratings started paying off

Once tourists could check reviews freely, demand moved toward better-rated places. Among restaurants that tourists could easily reach, revenue rose by about 5% overall, roughly 28,500 euros a year for an average restaurant. High-rated restaurants gained 5.6%, and 8% in the first year alone. Mid-rated ones gained 3.6% and grew their staff by 10%. Low-rated restaurants gained nothing.

High-rated restaurants added little staff, which suggests they were already full. Their revenue per employee rose by 4.3%, a sign that some of them raised prices once reviews vouched for their quality.

Location mattered in a surprising way. The well-rated restaurants that gained most were not those facing the monuments but those tucked a short walk away, the kind tourists find by checking a map and reviews on their phone. Restaurants in the most visible spots saw no clear gain.

Tourist traps closed, survivors upgraded

For poorly rated restaurants, the monthly chance of closing doubled after the change. Mid- and high-rated restaurants were no more likely to close. Counting new openings as well, in the most touristy neighbourhoods the share of low-rated restaurants fell by 3.75 percentage points while the share of high-rated ones rose by 4.6 points.

The low-rated restaurants that stayed open invested in quality, and the evidence comes from payroll records, not just reviews. They became 8% more likely to hire someone with restaurant experience (16% for mid-rated ones) and less likely to hire newcomers, and their daily wages rose by about 1 euro, or 2%. Their ratings then improved by about 0.1 stars, even when reviews from EU tourists were left out of the count. High-rated restaurants changed little.

What it adds up to

Scaling the results to Italy as a whole, Donati estimates that giving every customer cheap access to reviews would raise restaurant revenue by about 1.5% and employment by about 1.4%, around 174 million euros and 13,400 jobs, while raising closures by 0.4 percentage points a year. These are rough, back-of-the-envelope figures based on several assumptions, but the direction is clear: easier access to reviews made the whole market more quality-driven.

Beyond the restaurant table

Restaurants are a classic experience good: you only know the quality after paying. So are hotels, tradespeople, clinics, salons, holiday rentals, courses and many business services. My reading, not tested in the paper, is that the same pattern is likely wherever reviews suddenly become easier to see at the moment of choice, for example when maps, AI assistants and search results start summarising ratings for you. Good providers that were hard to find gain; weak providers that lived on location, passing trade or one-off customers lose their cover.

The study also shows what works for those on the wrong side of the ratings: not more marketing, but better people and better service, which then show up in the reviews.

What to expect

• Gains are modest but real. Revenue rose by about 4% to 6% for restaurants rated around or above 4 stars that tourists could reach.
• A short-lived boost at the top. High-rated restaurants gained 8% in the first year, but the longer-term gain was smaller and not statistically clear.
• Upgrading costs money. Low-rated survivors paid about 2% more in daily wages to hire experienced staff.
• Slow change in the market. Closures of low-rated restaurants picked up only about a year after the change, and the mix of restaurants kept shifting over time.

Where it does not work

• One city, one sector. The study covers restaurants in the province of Rome, a city full of one-time visitors, where the effect of reviews may be unusually strong.
• Only some customers were affected. Only EU tourists got cheaper data, so the estimates capture the effect of a partial shift, and the Italy-wide figures rely on several assumptions.
• Prices were not observed. The evidence that high-rated restaurants raised prices comes from revenue per employee and regional price data, so it is suggestive.
• Ratings are an imperfect measure of quality. Ratings may also have risen because different people were reviewing, though the gains for low-rated restaurants held when EU reviews were left out.

Questions marketers ask

Also new this week

What the researchers did

On a screen you see the whole shelf. On a voice assistant you hear one product, then a question: "Do you want me to order this?" To hear anything else, you have to say no, and remember what came before. Alex Mari (American University of Bahrain), Andreina Mandelli (SDA Bocconi) and René Algesheimer (University of Zurich) wanted to know what this does to choice.

They built an Alexa shopping app that copied the voice, wording and flow of Amazon's own voice shopping. In one-to-one video sessions, each participant sat alone with an Amazon Echo and bought a pack of four AA batteries by searching for the category, then saying "yes" or "no" to each suggestion. The first suggestion was set by the researchers: Amazon's own brand, AmazonBasics, or Duracell, at either the cheapest or the most expensive price. Every later suggestion was drawn at random from 35 brands and five prices, all with identical product descriptions. Participants could keep what they bought.

Participants

484 people recruited through two Swiss universities, average age 25; 85% rarely or never used a voice assistant

Task

Buy one 4-pack of AA batteries by voice on an Amazon Echo, searching by category, not brand

First offer

AmazonBasics or Duracell, at CHF 4.95 or CHF 6.95 (121 people in each of the four groups)

Measured

How many options each person heard, what they bought, decision satisfaction, and willingness to let the assistant choose and to trust its suggestions

The first option wins

Half of the shoppers (50%) said yes to the very first product, and 81% bought one of the first three. On average people heard about two and a half options before buying.

The brand placed first did most of the selling. When AmazonBasics led, 54.5% bought AmazonBasics; when Duracell led, 61.2% bought Duracell. Each brand was bought by only about 2% of people when the other one went first. Price behaved the same way: when the first offer was the cheapest price, 71% paid the cheapest price; when it was the premium price, 61% paid the premium price.

Brand equity still showed. Duracell, which more participants already knew, kept more buyers when it came first than the store brand did (61% against 55%).

Fewer options, happier shoppers

Research on screen-based shopping often finds that a very short list can leave people less satisfied with their choice. Here the pattern ran the other way. The more options people heard, the less satisfied they were with their decision, and the less willing they were to let the assistant decide for them or to trust its suggestions next time. Those who accepted the first offer were the most satisfied. Satisfaction explained only a small part of the link to trust and delegation.

But nobody should read this as consumers handing over the keys. On a 1 to 7 scale, willingness to let Alexa choose averaged 2.71 and trust in its suggestions 3.29, both below the midpoint. Decision satisfaction sat just above it, at 4.06.

Brand, price and the retailer's own label

Whether the first offer was the store brand or Duracell, cheap or premium, made no clear difference to how satisfied people were or how willing they were to rely on the assistant. One pattern did stand out: people were more willing to let the assistant choose when brand and price matched what they would expect, Duracell at the premium price or AmazonBasics at the low price. The authors suggest a mismatch, such as a cheap leading brand, may look suspicious.

For the retailer, leading with its own label paid slightly: shoppers paid CHF 5.92 per pack on average when AmazonBasics came first, against CHF 5.81 when Duracell did. The authors' explanation is that people who turned down the less familiar store brand then settled for the next good-enough option rather than hunting for a bargain.

Beyond the smart speaker

This study is about batteries on Alexa, but the set-up it tests is spreading: an assistant that offers one answer at a time and waits for a yes. AI chat assistants that recommend "the best option", shopping agents built into retail apps, in-car assistants, and AI search summaries that name one product all work this way. My reading, not tested in the paper, is that the same logic is likely to apply wherever choices are low-stakes and alternatives are hard to compare: the first suggestion does most of the selling, and the brand that is not in it barely gets heard.

That shifts the marketing question from "how do we stand out on the shelf?" to "how do we become the answer?" for brands, and to "how do we use that power without losing trust?" for whoever runs the assistant.

What to expect

• A strong first-offer effect for everyday items. About half of buyers took the first product and four in five bought one of the first three.
• Defaults steer price as well as brand. When the first offer was cheap, most people paid the cheapest price; when it was premium, most paid the premium price.
• Small revenue gains, not windfalls. Leading with the store brand raised the average price paid by CHF 0.11 per pack.
• Slow growth in trust. Satisfied shoppers were more open to delegating, but willingness to let the assistant choose stayed below the midpoint.

Where it does not work

• Association, not proof of cause. The researchers set the first offer, but people chose how many options to hear, so the link between fewer options and satisfaction may partly reflect who stops early.
• One low-stakes product. Batteries are cheap and similar. For expensive, emotional or complex purchases, hearing only one option may well leave people less satisfied.
• First-time, unpersonalised use. Most participants were young, new to voice shopping and could not ask for a brand by name, and the assistant knew nothing about their past purchases.
• No comparison with screens. The study did not include a website, a chat interface or another assistant, so it cannot say how much of the effect is due to voice itself.

Questions marketers ask

Also new this week

What the researchers did

Most social media ads, and most academic studies of them, are tested as a single post: a screenshot shown on its own. But people meet ads in a feed, scrolling past friends' updates, memes, news and other brands' ads. Platform A/B tests are no clean alternative either: the platforms' algorithms show each ad version to different kinds of people, which the authors call divergent delivery, so the comparison is no longer random.

Hauke Roggenkamp (ETH Zurich), Johannes Boegershausen (Rotterdam School of Management) and Christian Hildebrand (University of St. Gallen) built DICE, Digital In-Context Experiments: a free, open-source app that recreates a scrollable feed in the style of X or LinkedIn. Researchers or brands decide exactly which posts appear and in what order, and the app records how long each post stays on screen (dwell time), plus likes and comments, before handing people on to a survey. They demonstrated it with two studies and interviewed nine marketing managers.

Tool

DICE, a free, open-source app with a no-code setup (dice-app.org); feeds are built from a spreadsheet of posts

Study 1

982 US adults, a feed with an airline ad among ordinary or disaster-related posts

Study 2

300 US adults under 36, a 40-post feed of memes with five tech-brand ads in random positions

Measures

Dwell time per post, likes and comments in the feed, brand attitude, and brand recall

Interviews

Nine managers from consumer goods, apparel, food delivery and agencies

Bad neighbours hurt your brand

Brand safety is the risk of ads appearing next to content that damages the brand, something industry reports say up to 75% of brands have experienced. Testing it on a real platform would risk the very damage it measures.

In the first study, the same sponsored post from the airline KLM, promoting flights to Brazil, sat fifth in a feed. For half the participants the other posts were ordinary content about Brazil; for the other half they were real posts about the severe flooding in Brazil at the time. People who saw the ad among the flood posts rated KLM lower: 4.31 against 4.82 out of 7.

Attention made the difference. The more time people spent on the ad, the bigger the penalty from the unsafe context. For people who looked at it for less than about 2.3 seconds, the context had no detectable effect.

Attention, not position, drives recall

In the second study, five real ads from Apple, Bose, Nintendo, Samsung and Whoop were mixed at random into a feed of 35 memes. Ads placed higher in the feed were more likely to be recalled afterwards. But the reason was attention: ads higher up were looked at for longer, and once the time spent on each ad was taken into account, position no longer predicted recall. Time on the ad explained about half of the position effect.

Clicks and reactions told much less. Without any prompt to engage, fewer than half of participants liked any post at all, about one in ten liked an ad, and only four people out of 300 replied to one.

What managers saw in it

The managers interviewed saw three uses: more transparent insight than platform A/B tests, which "always keep some information to themselves"; a safe sandbox to pretest high-stakes campaigns before spending on media; and a way to compare the brand's own posts against strong influencer and user-generated content. They also valued the behavioural data as a complement to traditional ad copy surveys.

What to expect

• Context effects are real but moderate. An unsafe context lowered brand attitude by half a point on a 7-point scale, a meaningful but not dramatic effect.
• Smaller effects than in single-post tests. The authors warn that realistic feeds produce smaller, noisier effects, so feed tests need larger samples.
• Attention as the key metric. Time on the ad predicted recall better than where it appeared in the feed.
• Low cost. The app is free and open source; the main cost is paying a panel of respondents.

Where it does not work

• Not people's own feeds. Participants scroll a feed built by the researcher and know they are in a study, without personalised algorithms or real friends.
• Demonstrations, not definitive tests. The two case studies were designed to show what the tool can do, not to settle questions about brand safety or ad position.
• Dwell time is a proxy. Very short dwell times reliably signal inattention, but longer ones do not always mean more attention.
• One platform style. The studies used an X-style feed with US participants; results may differ on video-first platforms such as TikTok or Instagram Reels.

Questions marketers ask

Also new this week

What the researchers did

AI already negotiates for companies. Walmart uses AI agents to agree terms with suppliers, start-ups sell bots for sales, procurement and salary talks, and AI agents handle the first rounds of brand deals for influencers. Most are judged on the money they bring in. Sumon Chaudhuri (WU Vienna) and Arnaud De Bruyn (ESSEC) asked what happens to the person on the other side.

They trained three bots on 6,432 recorded games of a fast bargaining task, in which two players have ten seconds to agree how to split a sum of $1 to $6 and only the bot knows the true amount, much as a seller knows its costs better than a buyer. A human-like bot copied people's behaviour, quirks included. An efficient bot learned to maximise its earnings. A "superhuman" bot was trained to earn as much as possible while a second AI, acting as a judge, penalised anything that did not look human. Then 1,019 people played 20 games each against a bot or another person, with the bot shown either as a person or as "BargainBot".

Experiment

1,019 adults on Prolific, 20 incentivised games each; 15,285 games analysed after practice rounds

Conditions

Human-like, superhuman (three settings) or efficient bot, each shown with a human or a robot avatar, plus human-versus-human games

Measures

Agreements and earnings, a Turing test (bot or human?), and feelings about the outcome, oneself, the process and the relationship

Training data

6,432 games played by 110 students in an earlier study (Camerer, Nave and Smith, 2019)

What they found

The superhuman bot was the best negotiator. Shown as a person, it kept 39% of the pie per game, counting deals that fell through, against 33% for people in the same role and 36% for the bot that simply copied humans. It also reached more agreements than the copycat bot (67% against 58%), so its opponents earned more too.

The efficient bot, which looked best in simulations, disappointed. It made rapid-fire offers, replied in a fraction of a second and reliably accepted almost anything as the deadline approached, because any deal beats no deal. People spotted the pattern and exploited it, earning 34% of the pie against it and leaving it with 32%, no more than a human would have made. The more rational the bot, the more often players took more than half the pie, and even more so when told it was a bot.

More human than human

Players rarely spotted the better bots. Asked whether their counterpart was a bot or a person (1 = definitely a bot, 7 = definitely human), they rated the superhuman bot 4.11, higher than real human opponents (3.40), while the efficient bot's mechanical tactics gave it away (2.93).

Its tactics were human but ruthless. It would often open with nothing and wait for the other side to make offers, and sometimes let the clock run out rather than concede, like a tough human negotiator. "In negotiation, the first one to talk loses." The authors admit they lose to it regularly, despite knowing how it works.

The relationship cost

Winning the game did not win people over. Every bot, including the ones people thought were human, left them feeling worse about the outcome, the process and above all the relationship than negotiating with another person. Feelings about the relationship averaged 4.95 out of 7 after human opponents, but 3.21 to 3.90 after bots.

A human avatar and name did not help: people felt the same whether the bot was presented as a person or as "BargainBot". The authors suspect an uncanny-valley effect. People build a picture of their counterpart over repeated rounds and signal their intentions; the bots, trained one game at a time, ignored those signals, which may feel like talking to someone who never listens. Because such feelings predict how future negotiations go, the authors warn that AI negotiators could quietly erode customer and supplier relationships.

What to expect

• Better deals per negotiation. The best bot earned about 18% to 21% more than people playing the same role.
• More agreements. A well-designed bot closed more deals than a bot copying human behaviour (67% against 58%).
• Lower satisfaction. Expect counterparts to feel less positive about the relationship after dealing with any bot, by roughly one to almost two points on a 7-point scale.
• Learning by counterparts. Over repeated negotiations, people learn and exploit predictable bot tactics, especially once they know it is a bot.

Where it does not work

• A simple, low-stakes game. Ten-second bargaining over a few dollars is far from a contract or supplier negotiation; results may differ when stakes are high and talks involve words.
• Mismatched training data. The bots learned from US economics students, who were greedier than the general public they played against.
• One game at a time. The bots did not remember earlier rounds with the same person, which the authors suspect drives the relationship cost.
• No real relationships. Players did not have a lasting business relationship with the bot, so long-term effects are inferred, not measured.

Questions marketers ask

Also new this week