Is ChatGPT a real traffic source for online shops? What a Marketing Science study found
A year after ChatGPT began linking to online shops, researchers followed its visitors through 973 e-commerce sites. The channel is tiny, and its shoppers convert better than paid social but worse than almost everything else.
Across 973 online shops and $20 billion in sales, ChatGPT referrals made up under 0.2% of traffic and converted below most channels.
Marketing in 3 · Issue 33 · Week of 24 August 2026
Where a shopper asks matters as much as what they ask. This study tracks the ones who start with a chatbot. Photo: Jakub Żerdzicki on Unsplash.
The short answer
Not yet. A year after launch, visitors sent by ChatGPT were fewer than 0.2% of traffic. They converted more often than paid social visitors but less often than every other channel, and they spent less per visit. The picture is better for complex products and is improving over time. For most shops it is a small channel worth watching rather than a new engine for sales.
Key takeaways
Across 973 sites, ChatGPT sent under 0.2% of traffic, about 200 times less than Google's organic search.
Compared with ChatGPT visitors, organic search visitors were 13% more likely to buy and affiliate visitors 86% more likely. Only paid social did worse.
In complex categories such as finance, vehicles and business services, ChatGPT's share of traffic was 4.6 times higher and its conversion rate beat several traditional channels.
Over the year, conversion rose while order values fell, so revenue per visit improved only modestly.
What the researchers did
In August 2024, ChatGPT began adding links to online shops when a user showed they wanted to buy something. Ask for an
espresso machine and you get a comparison and buy links for your country. These links are organic, not paid for
by advertisers. Marketers called it a new channel before anyone knew how well it worked.
Maximilian Kaiser, at the University of Hamburg and the analytics firm Grips Intelligence, and Christian Schulze, at
Frankfurt School of Finance & Management, got inside the analytics accounts of 973 online shops. They could see which
channel brought each visitor and what the visitor did next, for a full year.
They compared ChatGPT referrals with eight established channels: paid search, organic search, direct, email,
referral sites, affiliates, paid social and a catch-all "other". They looked at money (how many visits end in a
purchase, how large the order is, how much a visit earns) and at behaviour (bounces, time on site, pages viewed).
Period
August 2024 to July 2025, with the main analysis on the six months to early August 2025
Shops
973 e-commerce websites across 24 categories and every continent, with $20 billion in yearly revenue
Volume
More than 50,000 purchases from ChatGPT referrals, against 164 million from traditional channels
Measure
Last-click attribution: the channel of the final visit before purchase gets the credit
Method
Regression models that account for how few ChatGPT visits each shop receives in a week
A small channel that buys less often
Start with scale. ChatGPT supplied less than 0.2% of all visits in the data. Over the six month main window it sent
about 4 million visits, where paid search sent about 1.5 billion. More than nine in ten of the AI-driven visits came
from ChatGPT. Perplexity, Gemini, Copilot and the rest were negligible.
Next, quality. The raw numbers look bad for ChatGPT, but they mislead. Half of the weekly ChatGPT figures contained
zero purchases, so simple averages are a poor guide. After the authors accounted for how many visits sat behind
each number, and for differences between shops, devices and months, the gaps narrowed. ChatGPT visitors still
converted less often than every channel except paid social.
The chart shows the odds of a visit ending in a purchase, with ChatGPT set to 100. Paid social comes in at 47,
about half as likely. Organic search is 13% higher. Affiliate visitors are nearly twice as likely.
Likelihood that a visit ends in a purchase, ChatGPT = 100
Likelihood that a visit ends in a purchase, ChatGPT = 100
Group
Value ()
Paid social
47
ChatGPT
100
Organic search
113
Referral sites
124
Email
132
Direct
133
Other
144
Paid search
145
Affiliate
186
Table 3, six month data. Regression coefficients on conversion, converted to relative likelihood against ChatGPT (e.g. +13% for organic search, +86% for affiliate, 53% lower for paid social).
People stay briefly but they do not leave at once
Engagement tells a mixed story. ChatGPT visitors bounced less than visitors from most channels. Email visitors were 20%
more likely to leave after one page, and organic search visitors 13% less likely. So the traffic is not junk, in the
sense of arriving by mistake.
Those visitors looked at fewer pages than every channel except paid social. Time on site landed in the middle.
Probably ChatGPT has already done the comparing, so the visitor arrives with a specific
product in mind and has less to browse.
Revenue per visit tells the same story as conversion. ChatGPT earned more per visit than paid social and less than
the other seven channels. Average order value was noisier, and for some channels the difference was not
statistically reliable.
Complex products fare better
The study then asked where ChatGPT works best. The authors sorted shops by how complex their products are. Finance,
vehicles, business and consumer services, law and government, and jobs counted as complex. News, sport and
entertainment counted as simple.
On complex sites, ChatGPT's share of traffic was 4.6 times higher. Its conversion rate also beat paid social,
referral, email, organic search and direct traffic, and its revenue per visit moved closer to search. On simple sites
the gap was widest. Their explanation is that a chatbot adds most when the buyer has a lot of information to
weigh and adds friction when the purchase is routine.
They also split shops by how tech-savvy their visitors are. Where visitors skew toward technology fans, ChatGPT's
traffic share was 3.8 times higher. Where visitors are younger, it was 5.5 times higher. In both groups the conversion
gap to other channels shrank.
This part of the study is exploratory. The split uses shop-level categories, not individual shoppers, so it points to
a likely explanation without proving it.
How many times higher ChatGPT's traffic share is
How many times higher ChatGPT's traffic share is
Group
Value ()
Complex vs simple products
4.6
Top vs bottom tech-savvy shops
3.8
Younger vs older visitor shops
5.5
Sections 6.1 and 6.2: complex against simple product categories, top against bottom quartile of technophile visitors, and younger against older visitors.
Conversion up, order values down
The channel changed during the year. Conversion at ChatGPT visits climbed steadily, while conversion at other channels
held steady apart from holiday peaks. Order values moved the other way. Average order value from ChatGPT drifted
down, which offset much of the conversion gain.
The net result was a modest rise in revenue per visit. The authors projected the trend forward. In most scenarios,
ChatGPT's conversion rate gets close to organic search within another year without matching it. Revenue per visit matches organic search only
under the most optimistic projection. They stress that projections are uncertain.
The same pairing, higher conversion and lower order value, shows up in the shops with the most tech-savvy visitors.
That is why the authors point to growing skill with chatbots as the likely driver: people who use them well buy more often,
and they buy cheaper things.
Early industry reports had pointed in opposite directions, from conversion rates several times higher than search to
rates lower than other channels. This study is the first of its size and sides with the cautious end.
Marketing in 3 steps
Ready. Set. Market.
1
ReadyDecide and prepare
Find your AI referrals. In your web analytics, open the report that lists visits by source. Filter for sources from AI chat tools such as ChatGPT. Write down visits, orders and revenue for the last 90 days.
Compare with two other channels. Make a table with three rows: AI chat, organic search and email. For each, list visits, conversion rate and revenue per visit. If AI chat has fewer than 30 orders, treat its rate as noise and wait another month.
2
SetBuild it
Check where those visitors land. List the ten pages that receive the most AI chat visits. Open each on your phone. Each should show a price, stock status and a clear buy button without scrolling.
Make complex products easy to compare. Pick your five products that need the most explaining. On each page add a short comparison table and a plain-language summary of who it suits. A chatbot can quote clear facts, and a shopper arrives ready to decide.
3
MarketLaunch and measure
Tag your own links. Where you can influence the link, such as a link you share in a chat or newsletter, add a tracking tag so the source shows separately in your reports. Check the next day that it appears.
Review monthly on three numbers. Each month, record conversion rate, average order value and revenue per visit for AI chat. If conversion rises while order value falls, check which products those buyers choose before you increase effort on the channel.
What to expect
Small numbers, so patience. Even a busy shop may see only a handful of AI-referred orders a week, which makes monthly figures swing.
Better results in considered purchases. Expect the strongest results where buyers compare many details, and weaker ones for routine buys.
Cheaper baskets for now. The trend in the data was more orders of lower value. Do not read a rising conversion rate as growth in revenue.
Credit shows up late. Last-click reports give ChatGPT credit only when it was the final visit, so earlier influence stays hidden.
Where it does not work
Descriptive, not causal. The study shows how channels compare. It cannot say that ChatGPT causes shoppers to buy less or more than they otherwise would.
Last-click credit. A shopper who researches in ChatGPT and buys later through search is counted as search. That understates the chat channel's role in discovery.
A young channel. The data stop in mid 2025. Instant checkout, shopping agents and paid ads in ChatGPT arrived afterwards and could change the picture.
Data from one firm. The data were supplied by Grips Intelligence, which employs the first author, and shops that share analytics may differ from those that do not.
Questions marketers ask
How much traffic does ChatGPT send to online shops?
Very little so far. Across 973 shops in a year, it supplied under 0.2% of visits, about 200 times less than Google's organic search (Kaiser & Schulze, 2026, Marketing Science).
Do ChatGPT visitors convert better than visitors from search?
Not on average. Organic search visitors were about 13% more likely to buy than ChatGPT visitors. Paid social was the only channel that did worse than ChatGPT.
Which products suit ChatGPT referrals?
Complex ones. In categories such as finance, vehicles and business services, ChatGPT's share of traffic was 4.6 times higher and its conversion rate beat several established channels.
Will ChatGPT referrals keep improving?
The trend was upward for conversion and downward for order value. Projections say conversion approaches organic search within a year, without a guarantee, and new features could change the path.
Why do industry reports disagree with this study?
Earlier reports covered far fewer sites and gave answers that ranged widely. This study covers 973 shops and adjusts for how few ChatGPT visits each shop receives, which makes simple averages unreliable.
An editorial that defines an AI agent by three features: autonomy, memory that adapts, and links to other tools. It separates advisory agents, which recommend, from transactional agents, which buy for you. Mapping six years of research, the authors find most studies look at reactions to recommendations and few look at how people prompt agents or how trust builds. Plan for a customer who may no longer make the final click.
Across three studies, a simple "trustworthy AI" certification label raised both thinking-based and feeling-based trust in an AI system and the intention to use it, with thinking-based trust carrying the effect. Labels that listed the principles behind the certificate did better, and seeing labelled systems lowered trust in unlabelled ones. If you want people to rely on an AI feature, show them why it can be trusted.
Source
Kaiser, M., & Schulze, C. (2026). Frontiers: ChatGPT Referrals to E-Commerce Websites: How Do LLMs Compare Against Traditional Channels?. Marketing Science, 45(4), 699-715. https://doi.org/10.1287/mksc.2025.0489
This brief is written for practitioners from the published article. Open access (CC BY 4.0): free to read at the publisher. Numbers are as reported by the authors; read the article for full methods.
Half of 484 voice shoppers bought the first product Alexa offered. What that means for brands, retailers and any AI assistant that recommends one option at a time.
Shoppers leaving a staff-free store pledged about 40% fewer volunteer hours than those leaving a staffed one. Why it happens, and three fixes.
Bruno Schivinski is Associate Professor of Marketing at Gdańsk University of Technology. He studies how consumers engage with brands in digital and social media.
