GEO works, but almost nothing being sold as GEO is the part that works. Two 2026 research efforts measured what actually gets a page cited by AI assistants. The strongest factor was whether the engine's own search retrieved the page at all, which explained 32% of the variation. Everything on the page put together explained 2.8%. The FAQ blocks, schema markup, summary boxes and author bios sold as "GEO packages" showed no measurable effect at all.
This article goes through what the studies did, the four findings that matter for your website, and what we changed in our own process because of them.
What the research actually did
Most GEO advice is written by people selling GEO. These two are not.
- A University of Toronto study (Chen and Koudas, October 2026) logged 158,099 AI answers containing 468,037 citations across ChatGPT, Gemini and Perplexity, every day for three months, then modelled which pages got cited and why. Crucially, it compared pages within the same website, which removes the usual confusion between "this tactic works" and "big sites happen to use this tactic".
- A critical survey of 45 GEO studies (Martinez, July 2026) reviewed everything published on the subject between 2023 and 2026. Its conclusion, in its own words: "no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior."
Read that second line again before you sign a GEO contract with a guaranteed outcome in it.
Finding one: being found beats being optimised
Before an AI assistant can cite your page, its own search has to retrieve the page. That step turned out to be nearly the whole game.
- Retrieval rank explained 32% of whether a page was cited. All page-level properties together explained 2.8%.
- Pages the engine ranked first were cited 43.8% of the time. Pages at ranks 11 to 20 were cited 3.5% of the time.
So the single highest-value thing you can do for AI visibility is ordinary, unglamorous technical and content SEO: a site that can be crawled, pages that are indexed, content that matches what people ask. Google says the same thing in its own documentation on AI features — a page needs to be "indexed and eligible to be shown in Google Search with a snippet", and "there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary."
One technical point deserves its own line, because it is the most common silent failure we find. AI crawlers do not run JavaScript. A log study of more than 500 million crawler requests found the major AI crawlers download JavaScript files and never execute them. Google and Bing render JavaScript; the assistants largely do not. If your main content is injected by a front-end framework, your page is effectively blank to them. Fetch your own page with a plain request and search the result for a sentence from the middle of your copy. If it is not there, nothing else on this list matters.
Finding two: FAQ blocks, schema and author boxes did nothing
This is the uncomfortable one, because it is what most GEO proposals consist of. Compared within the same website, here is what each tactic did to the likelihood of being cited:
| Tactic | Effect within the same site |
|---|---|
| FAQ block | −3.4%, not significant |
| Schema.org markup | −5.3%, not significant |
| Author information | −4.6%, not significant |
| Summary box | +6.9%, not significant |
| Page speed | −1.1%, not significant |
Compare different websites and the same tactics look great — FAQ blocks +18.1%, schema +7.5%, author information +14.9%. That gap is a textbook case of Simpson's paradox: well-resourced sites tend to have both FAQ blocks and the authority that actually earns citations, so the two get confused. The between-sites number is the one that ends up in sales decks.
None of this means you should rip out your schema. Structured data still earns rich results, still keeps your entity clear, and still helps a reader. Just do not buy it as an AI-visibility measure, and do not sell it as one.
Finding three: writing "for AI" can cost you
One benchmark in the survey took ordinary pages and rewrote their body copy specifically to be more citable. The rewrites went backwards:
- Presence in the top 20 retrieval results fell about 9%.
- Presence in the top 10 after reranking fell 16%.
- Final citations fell 6%.
The survey's phrase for this is that "citation-oriented rewrites can impair retrieval". It makes sense once you accept finding one: a page stuffed with answer blocks and hedged definitions is a worse match for a real query than the clear page it replaced, so it gets retrieved less, so it gets cited less. The same survey found keyword stuffing is now actively negative for AI citation, and that an "authoritative tone" had a weak and unstable effect.
If an AEO or GEO edit makes a page worse to read, it is making the page worse. That is the whole test.
Finding four: half the mentions have no page behind them
About 46% of the brand mentions in the study came with no supporting page cited at all. The model simply knew the name and used it.
You cannot reach that half with on-page work of any kind. It is reached by being written about, reviewed, compared, listed and discussed across the web — the same thing that has always built a brand, now with a second payoff. This is why our GEO work is mostly off your website: getting you into the roundups your buyers actually read, onto the review platforms your category uses, and into genuine conversations in the communities where your market talks.
It is also why ranking reports are a poor proxy here. In one analysis, 53% of the domains cited in Google's AI Overviews were outside the organic top 10, and 27% were outside the top 100 entirely.
So what does move the needle on the page?
One thing, measured repeatedly: matching the question in plain language. Within the same site, the levers that worked were
- the best-matching paragraph on the page — +15.5% per standard deviation,
- the title — +14.1%,
- plain word overlap with the question — +9.1%.
That is the entire on-page playbook: one page per real question, a title that names the question, and a first paragraph that answers it in the words your customer would use rather than the words your industry uses. No block structure required.
What we changed because of this
| Commonly sold as GEO | What the evidence supports |
|---|---|
| FAQ blocks and schema added for AI | Keep them for rich results and readers; expect no citation lift |
| An llms.txt file | One study of 137,210 domains found 97% of these files received zero requests. Five minutes of work, not a line item |
| Rewriting pages into answer blocks | Clear pages that answer one question; rewrites measured a 16% retrieval loss |
| Weekly AI citation reporting | Monthly, with a fixed prompt list. Month-to-month overlap of cited pages was 18%, against 45% for organic — weekly reporting is mostly noise |
| On-page work as the GEO deliverable | Technical health, question-matched pages, and third-party presence |
How to measure it without fooling yourself
Write down 25 to 30 questions a real buyer would actually type. Run them in ChatGPT, Gemini, Perplexity and Google's AI Mode. Record whether you were mentioned, whether you were cited, who was cited instead and which page they were cited from. Repeat monthly with the identical list.
Two warnings. Changing the prompt list destroys the comparison, so fix it on day one. And treat each assistant as its own channel: in one study of 22.7 million citations, ChatGPT and Perplexity shared only 10.9% of their sources for the same question. A win on one proves nothing about the others.
Is GEO worth paying for?
Yes, but not as a separate bolt-on product. The research says the work that earns AI citations is technical health, pages that clearly answer real questions, and presence across third-party sites — which is SEO, digital PR and content done properly. Be suspicious of any GEO package whose deliverables are FAQ blocks, schema and an llms.txt file, because those three measured no effect on citations within a site.
Does schema markup help with AI Overviews?
Not for citations. Google's own documentation states that there is no special schema.org structured data needed to appear in AI Overviews or AI Mode, and the Toronto study measured no significant within-site effect from schema. Keep your structured data for rich results, entity clarity and accurate machine-readable facts — those are real benefits — but do not expect it to get you quoted.
Why does my site not appear in ChatGPT or Perplexity?
The most common cause we find is technical: the content is rendered by JavaScript, and AI crawlers do not execute it, so they see an empty page. The second is that the search those assistants run does not surface the page at all, since retrieval rank explained 32% of citations in the research. The third is that your brand is simply not discussed anywhere else on the web, which no amount of on-page work can fix.
How long does GEO take to show results?
Expect to measure in months, not weeks, and expect noise. Month-to-month overlap of the pages cited in AI Overviews was 18%, compared with 45% for ordinary organic results, so a single week's movement tells you very little. Fix the prompt list you test with on day one and review it monthly.
Should I block AI crawlers?
It depends which ones. Search crawlers and training crawlers are separate decisions, and blocking the search crawlers removes you from those AI answers entirely. Blocking the training crawlers mainly opts you out of model training. If you want AI visibility, the search crawlers have to be allowed — and it is worth checking your CDN or firewall as well as your robots.txt, because one can block what the other permits.
If you want this checked on your own site, the first two things worth looking at cost nothing: whether your content survives with JavaScript switched off, and whether your robots.txt and CDN are quietly blocking the assistants you want to appear in. We are happy to look at both.
