Does personalization work, or does it creep people out? What a Journal of Marketing meta-analysis found
The privacy backlash everyone warns about is hard to find in 290 studies. What the data does show is a trap: a little personalization is the weakest position, and the best results sit at the far end.
A meta-analysis of 1,536 results from 290 studies finds personalization lifts responses with almost no backlash. Moderate levels do worst.
Threading the needle: the message has to be stitched to the person, and half a stitch holds nothing. Photo: Immo Wegmann on Unsplash.
The short answer
Yes, and the backlash is smaller than most marketers fear. A Journal of Marketing meta-analysis pooling 1,536 results found that personalization reliably improves how people respond (average correlation .23), while privacy worry, irritation and ad avoidance did not measurably rise. Data tied to who people are, such as interests and values, beats age, location and browsing history. And the curve is J-shaped: the weakest results sit in the middle, not at the top.
Key takeaways
Across 290 studies, personalization raised positive responses (average correlation .23) and left negative ones at zero (.02).
Stated intentions rose (.22). Measured behaviour on its own, such as clicks and orders, averaged .02, which cannot be told apart from zero.
Personalizing on age, location or past behaviour weakened the effect. Personalizing on interests and values did not.
A curve fitted the data better than a straight line in 9 of 10 tests: the weakest results sat in the middle of the personalization scale.
What the researchers did
Ask whether personalization works and you get two stories. In one, a message that speaks to you feels relevant
and gets a better hearing. In the other, it feels like being watched and you push back. The field calls this
the personalization paradox, and individual studies have found support for both sides.
Martin Eisend, Dominika Niewiadomska and Guda van Noort did what a meta-analysis does: they gathered the
studies and pooled the results. They searched four databases, followed the citations of every relevant paper
and also took in working papers, theses and conference papers, so that studies with dull results were not
left out.
A study counted if it compared a personalized message with a plain one, or asked people how personalized a
message felt. Each result was turned into a correlation. Responses were sorted into two groups. Success means
people liked the ad, trusted the brand, clicked or said they would buy. Failure means they avoided the ad,
grew suspicious, felt it was intrusive or worried about privacy.
The team then coded the studies by what data the message used, where it ran, who saw it and how it was shown.
They also fitted curves to see how the effect changes as personalization goes up. Last, they surveyed 198
marketing professionals to find out what practitioners believe.
Scope
1,536 effect sizes from 290 studies in 229 papers
Search
Papers and unpublished work up to 2024, across EBSCO, ScienceDirect, ProQuest and Google Scholar
Measure
Correlation between personalization and each response
198 marketing professionals, asked what they believe and what they do
It works, and the backlash barely shows
Take the success side first. Across 937 results, the average link between personalization and a positive
response was .23. That is a medium effect, big enough to see in a campaign. The strongest gains were in
credibility, loyalty, positive emotion, the value people see in an offer and willingness to pass it on. Trust
moved less.
Now the failure side, where the paradox says the damage should be. Privacy concerns, intrusiveness, avoidance
and suspicion all sat within a hair of zero, and none was statistically different from it. One measure did
move: negative emotion, at .14. But it rests on only eight results from five datasets, so I would not build a
policy on it.
The authors also checked whether backlash was quietly eating the gains. It was not. Whatever backlash existed
accounted for between 2.6% and 3.5% of the total effect.
So where does the fear come from? Partly from studies that did find backlash, which are real and get quoted.
The pooled picture says they are the exception, and that in the average study the ad worked.
How strongly personalization lifts each response (correlation, 0 to 1)
How strongly personalization lifts each response (correlation, 0 to 1)
Group
Value (r)
Credibility
0.51
Loyalty
0.43
Positive emotion
0.4
Value of the offer
0.39
Word of mouth
0.38
Trust
0.23
Purchase intention
0.22
Actual clicks and purchases
0.02
Eisend et al. (2026), Table 6, adjusted correlations. The last bar was not statistically different from zero.
Better than advertising in general
An earlier meta-analysis of advertising as a whole, co-written by Eisend, gives a yardstick. Against it,
personalized messages did better on every measure the two share: 52% stronger on attitudes, 37% on behaviour,
45% on emotion, and 175% and 179% on cognitive responses and credibility.
Those percentages sound large because the baseline is modest. A correlation of .33 on attitudes means
personalization explains about 11% of the variation in how people feel about the ad. For credibility it is
about 26%.
Now the part that does not match the headline. When the researchers looked only at what people actually did
(clicks, opens, choices, purchases), the average effect was .02 across 150 results, and that is
indistinguishable from zero. Purchase intentions, the answer to a survey question, were up at .22.
Read that as a warning about measurement. Personalization reliably changes what people think and say. Whether it
changes what they do depends on the setting, and the pooled studies cannot tell you which.
Personalized messages against advertising in general (correlation, 0 to 1)
Personalized messages against advertising in general (correlation, 0 to 1)
Group
Value (r)
Attitude, personalized
0.33
Attitude, advertising in general
0.21
Behaviour, personalized
0.27
Behaviour, advertising in general
0.2
Credibility, personalized
0.51
Credibility, advertising in general
0.18
Eisend et al. (2026), Table 9. Weighted as in the earlier advertising meta-analysis, so the figures differ from the paper's main table.
What you personalize with decides the result
The coders sorted the data behind each message into types: demographic (age, gender), geographic (location),
behavioural (browsing and purchase history), socioeconomic (income, social ties) and psychographic
(personality, values, interests).
Personalizing on demographic, geographic or behavioural data produced weaker effects. Psychographic data did
not. The authors' reading is that this kind of data is tied to the customer's sense of self, so the message feels
like it is about them. Age, location and last week's clicks describe a person without being them.
How you got the data matters too. Messages built on data inferred from behaviour did worse than messages built
on what customers stated. Effects were stronger when people actively handed over the information, and
backlash fell when they were told their data was being collected and used.
Then what did not matter. Channel made no significant difference, whether the message ran on social media or elsewhere, online or
offline. Gender, age and student status did not either. That goes against the common belief that
older customers push back harder.
One setup effect deserves a flag. Studies in which participants merely imagined a personalized message
showed stronger effects than studies that showed a real one. Hypothetical scenarios flatter personalization.
A J, not a hill
The best known model of the paradox is a hill. A little personalization helps, a lot creeps people out, so
aim for the middle. Five competing curves were set out in the paper, and the team tested which one the data
fit.
The hill lost. The data fit a J. Plotted against how personalized the message was, success results dipped as
personalization rose from low to moderate, then climbed to their best at the high end. Failure results made the
mirror image. In nine of ten statistical tests, a curve fitted better than a straight line.
The authors' explanation is privacy cynicism. A little personalization feels creepy. A lot of it tells the
customer the data collection is everywhere and nothing can be done about it, and resignation lowers the
resistance. They did not measure that directly, so treat it as their best reading and not a finding.
The practical message in the paper is blunt: favour high levels of personalization and avoid moderate ones. My
own picture of the moderate zone is a message that knows your name and one product you looked at. The paper does not
define it that way, but that is the weakest place to stand.
Where marketers' habits and the evidence part ways
The survey of 198 practitioners shows how far habit sits from the evidence. About two thirds expected
positive effects and one third expected mixed ones. Most said they mainly use demographic, geographic and
behavioural data, mostly on social media and email. Many run moderate levels of personalization while believing
high levels work best.
Respondents also said they value transparency and underuse it. On a seven-point scale they rated the paper's findings
on data types and levels as surprising (3.96 to 4.95). They rated those same findings as useful (5.49 to 6.22).
It is a small survey of professionals recruited through an online panel, so take it as a sketch and not a census.
But the gap it points to, between the data most teams hold and the data that works best, is the practical core
of the paper.
Marketing in 3 steps
Ready. Set. Market.
1
ReadyDecide and prepare
List what you personalize with today. Write down every piece of customer data your messages use. Next to each one, write how you got it: the customer told you, or you worked it out from what they did.
Sort it into three piles. Pile one is who they are, like age, gender and city. Pile two is what they did, like pages viewed and past orders. Pile three is what they care about, like interests, goals and taste. Plan to lean on pile three.
2
SetBuild it
Ask for pile three. Add one question at sign up or at the first order: "What are you shopping for?" Give four or five answers to pick from and let them choose up to three. Use those answers in the next message you send.
Build three versions of one message. Version A is the same for everyone. Version B adds a name and one thing they looked at. Version C uses what they told you they care about and reads as if it was written for one person. Make C feel personal, not like a template.
3
MarketLaunch and measure
Tell them why they are seeing it. Add one plain line at the bottom of the personal versions: "You are seeing this because you told us you like hiking." Say where the information came from, in words a friend would use.
Test for seven days, then count orders. Send A, B and C on the same day to three groups of equal size. After seven days, compare orders, clicks, unsubscribes and spam complaints for each. Judge by orders per 1,000 people. If C beats B, make C your default. Do not settle on B.
What to expect
Bigger lifts in what people say than in what they do. Opinions and intentions move easily. Clicks and orders move less, so judge your test on orders and not on how people say they feel.
The middle is the expensive trap. Half-hearted personalization costs the same data work as the real thing and lands in the weakest part of the curve. Either commit or keep the message plain.
Published averages are a ceiling, not a promise. Effects were smaller in newer studies and smaller again in unpublished ones, so plan for less than the headline number.
Your own result will differ. The studies disagreed with each other a great deal, and the statistical tests for that disagreement were highly significant. Your list, product and channel will land somewhere in the spread.
Where it does not work
Averages across very different studies. Products, channels, scenarios and measures varied widely. An average tells you where to start testing, not what will happen to you.
Correlation, not cause. The moderator results show which kinds of studies found bigger effects. The authors say plainly that this cannot prove that the data type itself caused the difference.
The J rests on a subset. Under four in ten studies reported how personalized their message was. Only 335 of the 943 success results and 71 of the 188 failure results could be used to draw the curve.
Publication bias is not ruled out for the success results. The funnel plot checks flagged asymmetry, and unpublished studies showed smaller effects. Studies run to 2024, so newer AI-driven personalization is barely covered.
Questions marketers ask
Does personalization work in marketing?
On average, yes. Pooled studies show positive effects on attitudes, credibility, loyalty and purchase intentions, with an average correlation of .23. Measured behaviour was the weak spot, close to zero, so test on orders in your own setting (Eisend, Niewiadomska & van Noort, 2026, Journal of Marketing).
Does personalization cause privacy backlash?
Barely, in the pooled data. The measures of worry, irritation, avoidance and distrust were all statistically indistinguishable from zero, with one small exception on negative feelings. Individual studies do find backlash, but it is not what the average looks like.
What kind of customer data works best for personalization?
Data tied to who the customer is: interests, values and personality. Age, location and browsing history gave weaker results. Information customers gave you directly beat information you inferred from their behaviour, and telling them you use it reduced backlash.
How much personalization is too much?
The evidence points at the middle, not the top. Results follow a J shape: persuasion is weakest at moderate levels and strongest at the highest. The authors suggest that heavy personalization produces resignation about data collection, which lowers resistance, though they did not test that directly.
Do older customers or certain channels respond differently to personalization?
Not in this analysis. Neither the channel nor the audience's gender, age or student status changed how well personalization worked significantly, and older audiences were not more resistant.
An online experiment with 1,256 people varied how sensitive the data behind an ad was and how many types of data it used. Ads built on sensitive data, or on several kinds of data, made people feel more watched. Adding more data on top of sensitive data did not make it worse. Feeling watched lowered attitudes to the ad, the brand and the platform, but did not raise privacy resignation. If you must use sensitive data, the amount matters less than the fact you used it.
Differential privacy adds controlled noise so a targeting decision cannot reveal who is behind it. A field experiment with 747,975 customers shows the placement matters. Applying the protection at the moment of the targeting decision earned about five times the profit of the alternatives, and both approaches stayed profitable under strong protection. If your team is adding privacy protection to targeting, ask where in the workflow it goes before asking how much to add.
Source
Eisend, M., Niewiadomska, D., & van Noort, G. (2026). Personalization in Marketing Communication: A Meta-Analysis. Journal of Marketing. https://doi.org/10.1177/00222429261460888
This brief is written for practitioners from the published article. Open access (CC BY-NC 4.0): free to read at the publisher. Numbers are as reported by the authors; read the article for full methods.
A "recommended for you" weight-loss refill lowered ratings of the pharmacy from 3.79 to 3.02 out of 7. When personalization embarrasses customers, and what to do.
In the Ferragni and Balocco case, 96% of journalists' tweets were about the influencer. Consumers followed the press, and the brand barely got mentioned.
Facebook reported a return of 7.47 on retargeting. A field experiment across 210 markets found the retargeting budget could be cut to zero without losing a single order.
Bruno Schivinski is Associate Professor of Marketing at Gdańsk University of Technology. He studies how consumers engage with brands in digital and social media.
