Key Takeaways
- A Frankfurt School study of 973 e-commerce sites (over USD 20 billion in combined revenue) over one year found ChatGPT referrals account for less than 0.2% of traffic, with conversion rates below most traditional channels
- The averages hide a sharp split by category: ChatGPT outperforms traditional channels for complex purchases like vehicles and finance, while its value stays limited for simple products
- Conversion rates rose steadily across the study period, and features like Instant Checkout and agentic shopping could change the picture in the years ahead
The 'Google killer' headline versus the empirical data
ChatGPT was quickly framed as a potential Google killer in online shopping. A new study by Frankfurt School of Finance & Management paints a more nuanced picture.
techxplore.comWhen ChatGPT began surfacing product recommendations, the narrative spread fast: this would replace Google Search and affiliate links alike. Among e-commerce operators, the prospect of AI assistants becoming the next major battleground stirred a mix of excitement and anxiety.
Against that heat, research by Professor Christian Schulze of Frankfurt School of Finance & Management and Maximilian Kaiser of the University of Hamburg offers some sober numbers. The paper, "ChatGPT Referrals to E-Commerce Websites: How Do LLMs Compare Against Traditional Channels?", is published in the journal Marketing Science and rests on real data from 973 e-commerce sites with combined revenue above USD 20 billion, spanning August 2024 to July 2025. The comparison set lines up the established channels: organic search, paid search, email marketing, affiliate marketing and paid social.
The conclusion, in a sentence, is that ChatGPT has not "already transformed" e-commerce. It is better described as a sprout that may yet grow into something significant. But that sprout is not growing evenly across every product.
By the numbers, ChatGPT is not the lead actor yet
The first thing the researchers showed is how small the traffic is. A full year after organic shopping links went live, LLM-driven traffic remained under 0.2% of the entire sample. That is roughly one two-hundredth of organic search, a volume that is easy to dismiss on size alone.
On quality, it still trails the established channels. ChatGPT's conversion rate and revenue per session beat paid social, but fall short of everything else. Coverage of the study notes it converted around 13% worse than organic search and about 86% worse than affiliate referrals. With more than 50,000 ChatGPT-driven purchases measured against 164 million transactions from traditional channels, this is hard to write off as a thin-sample fluke.
What is intriguing is that engagement looked relatively healthy. Metrics that signal interest, such as low bounce rates and session depth, were not bad, yet that interest rarely converted into a final purchase. Schulze explains the gap through trust: people do not use ChatGPT as the last step before buying, but check other sources first and then buy. The picture is of consumers treating the AI's suggestions as a reference opinion while completing checkout somewhere familiar.
The real story is in the product categories
Left there, this would simply read as a story about overhyped expectations. But the heart of the research starts here. As Schulze himself put it, the real story is in the product categories, and beneath the averages the results split sharply in two.
In complex, information-heavy categories that demand comparison and deliberation, ChatGPT already shows real substance. In areas such as vehicles, finance, insurance and business services, sites captured roughly 4.6x the LLM traffic share of simple-category sites, and beat five traditional channels: paid social, referral, email, organic search and direct. Where consumers want options structured, made comparable and supported through the decision itself, ChatGPT genuinely helps.
For quick-decision, low-involvement items such as news, sports and entertainment, the ChatGPT effect was negligible. That makes sense: when there is little to deliberate over, there is little reason to consult an AI, and searching and buying directly is faster. The researchers caution that such categories offer no justification for reshuffling lower-funnel marketing budgets.
| Dimension | Where it works (complex purchases) | Where it struggles (simple purchases) |
|---|---|---|
| Product category | Vehicles, finance, insurance, business services | News, sports, entertainment and other quick-decision items |
| ChatGPT referral share | Roughly 4.6x that of simple categories | Negligible, close to noise level |
| Vs. traditional channels | Beats paid social, referral, email, organic search and direct | Beats only paid social |
| How consumers use it | To compare, narrow down and support the decision | Buying directly without AI is faster |
| Implication for retailers | Worth capturing through information design and content optimization | Not a reason to reallocate lower-funnel budget |
The more complex the product, the more ChatGPT matters. Flipped around, that means the starting point for any retailer is to judge whether their own product is something consumers agonize over before buying.
How to pull 'complexity' toward your own catalog
So what should operators in information-intensive categories do? The key is to prepare, on your own content side, the kind of structured information ChatGPT is good at organizing.
ChatGPT can only help with comparison and filtering when the underlying information is already organized. Products whose decision criteria are stated explicitly in text, such as spec differences, fit by use case, and how to choose across price tiers, give the AI more to work with when it assembles a reason to recommend. A page built only from images and a short tagline tends to become, for an LLM, a product with nothing to compare. This is a slightly different mindset from classic SEO, an information design meant to be read and cited by AI, and it is worth pursuing alongside organic search optimization.
One caveat: the study does not say "move your budget now if you sell complex products." Traffic is still under 0.2%, and the main battleground remains the existing channels. Schulze advises companies to assess carefully whether their products are the kind that benefit from this guidance. Front-loading investment on excessive optimism is as poor a call as ignoring the trend entirely and failing to prepare.
Reconciling this with Adobe's findings
It helps to add one reference line here. Throughout 2025, Adobe Analytics repeatedly reported explosive year-over-year growth in AI-assistant traffic, with conversion rates above non-AI traffic. During the holiday season, conversions ran more than 30% higher than non-AI traffic, a figure that looks, at first glance, like the opposite of the Frankfurt School conclusion.
The two are not contradictory; they look at different cross-sections. Adobe's figures center on growth rates and relative comparisons, and a large growth rate is entirely compatible with a small base (under 0.2%). The measurement scope, timing and definition of AI traffic also differ. The peer-reviewed Frankfurt School work paints a conservative picture in which absolute volume is still small and lags traditional channels, while Adobe conveys a different truth: the momentum of the launch is real. For a business decision, the safe move is to keep both in mind, the scale of the momentum (Adobe) and the current size (Frankfurt School).
Conclusion
What the Frankfurt School research shows is that framing the ChatGPT-shopping debate as a binary of "works" or "does not work" is itself too crude. Traffic is still under 0.2% of the total, and conversion falls short of most traditional channels. Yet in complex categories like vehicles and finance, it already outperforms established channels. Only by looking category by category, rather than at the average, does the meaning for your own business come into focus.
The right move for e-commerce operators today is neither over-investment nor dismissal, but judging whether your products are the complex kind that consumers deliberate over, and if they are, putting your comparison and selection criteria into text. Conversion rates kept rising across the study period, and as the purchase experience improves through features like Instant Checkout and agentic shopping, this assessment could be rewritten within a few years. For now, this is the phase to do the unglamorous groundwork of information design, ahead of the main event to come.
