The short version
- A 2026 review of 45 studies by Martinez (opens in a new tab) concluded that claims about GEO return on investment clearly outstrip the academic evidence.
- In the clearest field test, by Watanabe and Nakayashiki (opens in a new tab), ChatGPT referrals to one site grew 5.7 times, but pages that got no optimization grew 3.5 times anyway.
- In a panel of US adults, Pew researchers (opens in a new tab) found just 1% of visits to pages with an AI Overview led to a click on a cited source.
- In an audit of real shopping questions, Uberti-Bona Marin and colleagues (opens in a new tab) found ChatGPT stated its own product preference in 79% of answers that recommended products.
- A 2026 framework by Kato and colleagues (opens in a new tab) shows how to value AI visibility like a media channel, but it has only been tested on simulated sales data.
Is there evidence that AI visibility leads to sales?
No published study has yet linked appearing in AI answers to sales or revenue. Martinez (opens in a new tab), at Sciences Po, reviewed 45 studies published between late 2023 and mid-2026. The section on business outcomes is titled “Traffic and Conversions: The Weakest Evidence.”
His verdict is blunt: at this stage, claims about GEO return on investment clearly outstrip the academic evidence. He lists the claim that citation scores predict clicks, conversions or revenue as rejected as a general claim. Only one suggestive field test and a few industry reports exist.
One of those industry reports describes a 20% production traffic lift against a control group. The review notes it does not describe group sizes, how pages were assigned or how uncertain the figure is. That is traffic, not sales, and it cannot support a general estimate.
What does the best real-world test show?
It shows a likely traffic gain, not a sales gain, and even the traffic gain is uncertain. Watanabe and Nakayashiki (opens in a new tab) work at Glasp, a web highlighting service, and studied their own site. They optimized one section of it and left the rest alone as a comparison.
Total ChatGPT referrals grew by a factor of 5.7, but untouched pages grew by a factor of 3.5 as ChatGPT itself grew. After allowing for that, they estimate an extra lift of about 1.82 times for the optimized pages. A stricter check found jumps almost as large before the work began, so the authors call the effect suggestive, not conclusive.
The study measured visits from ChatGPT, not sign-ups or revenue. We cover how to read results like these in our article on crediting ChatGPT referral growth and in how to prove GEO caused a change in sales.
Why isn’t website traffic a fair measure of the value?
Because much of what AI answers do for a brand happens without any click. In a panel of 900 US adults tracked in March 2025, Chapekis and colleagues (opens in a new tab) at Pew found just 1% of visits to pages with an AI Overview led to a click on a cited source. An AI Overview is the AI summary at the top of Google’s results.
Assistants show the same pattern. In a browsing panel studied by Iannelli and Ai (opens in a new tab), 34.1% of sessions that used an AI assistant showed no visit to any outside website. The authors work for Scrunch AI, a vendor of AI visibility tools, and note a session with no visit is not proof the question was answered. See our article on assistant sessions without website visits.
AI search can also make it harder to reach your site at all. In a one-week experiment with 1,100 US Google users, Wang and colleagues (opens in a new tab) found that switching people to AI Mode cut their click-through rate by 18.8 percentage points. Among participants who described their week with AI Mode, 15.3% mentioned difficulty getting to specific websites.
Can an AI answer influence buyers who never click?
Very likely, though no study has yet measured the purchases that follow. AI assistants increasingly act like advisers rather than lists of links. In an audit of real shopping questions, Uberti-Bona Marin and colleagues (opens in a new tab) found ChatGPT stated a first-person product preference in 79% of answers that recommended products, against 7% for Gemini and 2% for AI Overviews.
The same audit notes that, in the largest published measurement of ChatGPT use, product and service recommendations were roughly 2% of conversations. Small as a share, that is a large number of buying conversations. The authors also found that the recommended products often changed when the same question was asked again.
Buyers already lean on AI search. Among participants in a 2024 study by Li and Aral (opens in a new tab) at MIT, 85% used AI tools, and 63% of those used them to look up information. Their experiment with 4,927 Americans found that citations raised trust in AI answers even when the links were wrong; see our article on citations and trust.
Does every kind of appearance count the same?
No: being described when someone names you differs from being recommended when they don’t. In a study of 112 Product Hunt startups, Sharma (opens in a new tab) found ChatGPT recognized 99.4% of them when asked by name. Asked open discovery questions, it surfaced them only 3.32% of the time.
Appearances also come and go. In our consistency study, only 25.2% of the brands ChatGPT named for a buyer question appeared in all five runs of that question. A single appearance in one answer is a sample, not a position you hold.
And an appearance can carry the wrong details. In our study of business facts, 18.9% of 636 answers from four AI engines gave at least one address, phone number, website or opening time that differed from the Google profile. A mention that sends a buyer to an old phone number has negative value.
How should executives measure the ROI of generative search visibility?
Count how often buyers likely notice your brand in answers, then tie that to sales using a comparison group. Kato and colleagues (opens in a new tab) propose fitting AI visibility into marketing mix modeling, the method many firms already use to value media spend. Their key point is that a mention in an answer is not yet a media input.
To turn it into one, the authors combine four things:
| Input | What it means | Where it comes from |
|---|---|---|
| How often answers name you | Share of repeated answers to relevant questions that mention the brand | Running your buyer questions many times |
| How many such questions are asked | Volume of relevant questions in each market and period | Market estimates |
| Which AI tools handle them | Each assistant’s share of those questions | Usage data |
| Whether readers notice | Chance a reader registers the name in the answer | User studies |
They tested the idea with 2,240 answers to 56 product questions in English and Japanese. The brand they tracked was Glasp, the same site as the field test above; it appeared in 33.8% of answers from one GPT model and 27.8% from another. The link to sales in their paper is simulated, so the method is promising but unproven on real revenue.
The authors also stress that a content change can pay off outside AI answers, through ordinary search traffic or a better landing page. A fair measure counts those paths too, rather than crediting only the AI mention. Measurement like this is also the base of a longer-term AI search strategy.
What should you do about it?
Treat AI visibility as a channel worth testing, and do not book its return until you have measured it.
- Measure exposure, not just visits. Track how often each assistant names you across repeated runs of your buyers’ real questions; referral traffic misses most of the influence.
- Separate being described from being recommended. Report questions that name your brand apart from open category questions; the second is where new demand comes from.
- Check what the answers say. Audit prices, contact details and claims, because a wrong mention can cost you a customer.
- Build in a comparison group. Leave some comparable pages, products or markets untouched so platform growth is not mistaken for your work, as explained in our article on judging a rise in AI visibility.
- Connect to the models you already use. If finance runs marketing mix modeling, add noticed AI mentions as an input rather than building a separate scorecard.
- Plan for the click losses too. Value from AI answers sits beside traffic you may lose; see what lost clicks mean for pipeline and whether citations make up for them.
If you want help setting up this kind of measurement, see our generative engine optimization service.
What does the research not tell us yet?
The research has not yet answered the central question: how much revenue an AI mention is worth.
- No sales outcome has been measured. The best field test tracked ChatGPT visits on one site, and its effect is only suggestive.
- The valuation method is untested on real sales. The marketing mix framework uses real answers but simulated business results.
- How many people notice a mention is unknown. No published study yet measures how often readers register a brand name inside an answer.
- Influence without clicks is inferred, not counted. Studies show trust and stated preferences, not the purchases that follow.
- Most evidence is narrow. Key studies cover one site, US panels or a few weeks, and some come from AI visibility vendors.
Frequently asked questions
Does being cited by an AI search engine actually generate business value?
It can, but nobody has yet measured how much. A 2026 review of 45 studies found the claim that citations predict conversions or revenue is not supported as a general rule.
How should executives measure the ROI of generative search visibility?
Measure how often buyers are likely to see your brand in answers, then compare sales against a group that got no GEO work. A 2026 framework adds noticed AI mentions to marketing mix modeling, though it has only been tested on simulated sales.
Is AI referral traffic a good measure of GEO success?
Only partly. In a Pew panel, 1% of visits to pages with an AI Overview led to a click on a cited source, so most of the exposure never shows up as traffic.
Do AI recommendations influence what people buy?
They likely do, but purchases have not been measured. ChatGPT stated a personal product preference in 79% of answers that recommended products in one audit, and people trust answers more when they carry citations.
Sources
- Martinez (2026), Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026) (opens in a new tab), arXiv:2607.14035.
- Watanabe and Nakayashiki (2026), Disentangling Answer Engine Optimization from Platform Growth: A Log-Based Natural Experiment on ChatGPT Referral Traffic (opens in a new tab), arXiv:2606.04362.
- Chapekis, Lieb, Shah and Smith (2026), Investigating Click Behaviors On Google Search Result Pages That Produce an AI Overview (opens in a new tab), arXiv:2608.04831.
- Iannelli and Ai (2026), The New Shape of Search: How Conversational AI Recomposes Information Seeking (opens in a new tab), arXiv:2607.04282.
- Wang, Gleason, Bart, Wilson and Metaxa (2026), AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence (opens in a new tab), arXiv:2608.18352.
- Uberti-Bona Marin, Bertaglia, Astante, Rijsbosch, van Dijck, Hannák, Spanakis and Kollnig (2026), "If I Had to Buy Just ONE: Galaxy S26 Ultra": Auditing AI-Generated Product Recommendations (opens in a new tab), arXiv:2609.18729.
- Li and Aral (2025), Human Trust in AI Search: A Large-Scale Experiment (opens in a new tab), arXiv:2504.06435.
- Sharma (2026), The Discovery Gap: How Product Hunt Startups Vanish in LLM Organic Discovery Queries (opens in a new tab), arXiv:2601.00912.
- Kato, Honma and Kato (2026), Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact (opens in a new tab), arXiv:2609.11915.
- Underneath (2026), Ask an AI the same question 5 times: do the brands change?
- Underneath (2026), Do AI answers match a business’s Google profile?