Generative Engine Optimization Statistics, Source-Checked

Key points

• The "up to 40%" lift is relative. A source's share of a lab-written answer went from 19.4% to 27.3%, and the top method still lost on 27.6% of queries.
• Pew saw fewer clicks on Google searches that showed an AI summary, 8% of visits against 15%. That's an observed gap, not a tested cause.
• Engines lean on different sites. In Vector Index's sample of Cincinnati questions, ChatGPT cited bbb.org in 14 of 68 answers and Gemini cited it in none.

Which generative engine optimization statistics can you trust?

Statistic	What was counted	Who measured it, and when	What it doesn't show
"Up to 40%" more visibility	A source's share of answers that gpt-3.5-turbo-16k wrote from five supplied sources	The GEO paper's abstract (arXiv), accepted to KDD 2024. Re-derived by Vector Index from the authors' released results	Live engines, clicks, calls or sales
8% vs 15% of visits	Clicks on a traditional Google result, on searches with and without an AI summary	Pew Research Center, 900 U.S. adults, March 2025	Why the clicks differed, or anything about ChatGPT or Gemini
More than 60 percent incorrect	Eight generative search tools asked to identify news articles from excerpts	Tow Center, Columbia Journalism Review, tests run February 2025	Accuracy on questions about local businesses
bbb.org in 14 of 68 ChatGPT answers	Local answers that cited bbb.org at least once, through the API with web search on	Vector Index, August and September 2026	Other cities, the consumer apps, or a rate

What does the "up to 40%" GEO figure measure?

The abstract of the GEO paper by Aggarwal and colleagues says "GEO can boost visibility by up to 40% in generative engine responses." Vector Index re-derived that figure from the authors' released per-query results:

• Left unchanged, the chosen source averaged 19.4% of the answer across 989 queries.
• Rewritten with the top method, Quotation Addition, it averaged 27.3%. That's 7.9 percentage points, a 40.6% relative change.
• Quotation Addition beat the unchanged source on 64.8% of queries and lost to it on 27.6%.

Hypothetical example: a vendor tells you GEO lifts visibility by 40%. If your page made up 20% of an answer, a 40% relative lift takes it to 28%. That's 8 points on a lab measure, and nobody has checked yet whether a live engine retrieves your page at all.

The released code writes answers with gpt-3.5-turbo-16k from five sources supplied to it, and it doesn't query ChatGPT, Gemini, Claude, Perplexity or Google AI Overviews. Some released rewriting prompts also allowed invented material, including one that reads, "You may invent these sources but ensure they sound plausible and do not mislead the reader." If you borrow the lab's top methods, every quote and number you add to your own pages has to be real and sourced.

Do AI summaries on Google mean fewer clicks?

Pew found that Google users clicked less often on searches that showed an AI summary. It observed that pattern in the browsing data of 900 U.S. adults in March 2025. It didn't test whether the summary caused it.

• Users who saw an AI summary clicked a traditional result in 8% of visits.
• Users who didn't see one clicked a result in 15% of visits, nearly twice as often.
• A link inside the summary itself got a click in 1% of visits to pages with a summary.

The two groups of searches weren't alike to begin with. Summaries appeared on 53% of searches of 10 words or more and on 8% of one- or two-word searches. With searches that different, the click gap alone can't show cause and effect.

It still tells you where to look first. Long searches, and searches that open with "who" or "what," got summaries far more often in Pew's sample, so if your buyers search that way, check what Google shows them first.

Do ChatGPT, Gemini and Perplexity cite the same sites?

Not in Vector Index's sample. For the local AI citations study, Vector Index called ChatGPT (gpt-5.4), Gemini (gemini-3.6-flash) and Perplexity (sonar) through their APIs with web search enabled, on 24 and 25 August and 16 September 2026, and counted the domains cited in 168 answers to 38 question wordings that name Cincinnati or say "near."

• ChatGPT cited bbb.org in 14 of its 68 answers. Gemini cited it in none of its 68.
• Gemini cited reddit.com in 14. Perplexity cited reddit.com in 13 of 32.
• yelp.com was cited in 1 of 168 answers.

The engines searched differently too. Across all 48 question wordings in the same sample, ChatGPT ran 478 searches of its own across 78 answers, 6.1 per question, and Gemini ran 117 across its 78, 1.5 each (what AI searches for when asked to recommend a business).

These are readings on the days sampled, not rates. Every local question named Cincinnati, API answers aren't consumer-app answers, and Google AI Overviews, Claude and Microsoft Copilot weren't sampled. Still, a single "AI visibility" score averaged across engines can hide which sites each one leans on, so check your listings engine by engine.

How accurate are AI search engines at citing sources?

In one news test run in February 2025, the tools got most answers wrong overall. The Tow Center, publishing in Columbia Journalism Review, gave eight generative search tools excerpts from news articles and asked each to name the article's headline, publisher, date and URL. Collectively, they gave incorrect answers to more than 60 percent of queries, and Perplexity answered 37 percent incorrectly.

That test covered news, not local businesses, and the products have changed since. So when an answer cites a page about you or a competitor, open it and confirm it says what the answer claims.

How do you check a GEO statistic before you pay for work based on it?

1. Find who counted. A statistics page citing another statistics page isn't the source.
2. Name the measure. Share of an answer, clicks, citations and sales are different things, and a gain in one doesn't prove a gain in another.
3. Name the engine and how it was reached. A lab model, an API call and the consumer app can answer the same question differently.
4. Check the date. Engines change between samples.
5. Find the baseline and the losses. A relative lift needs its starting point, and an average hides the cases that went the other way.
6. Ask who wrote the questions. Vector Index wrote its own, not drawn from search logs, so its results describe those questions only.

If no figure passes for your business, take your own reading. Type a question a customer would ask into ChatGPT, Gemini and Perplexity. Save the engine name, the date and the full answer. Open every cited source and check whether your services and locations are described correctly.

Before you pay for GEO work justified by one number, read the up to 40% figure, set against its own baseline. You'll see the per-query results behind it and which methods lost to an unchanged page, so you know what evidence to ask for when a vendor quotes you generative engine optimization statistics.