What nobody can promise you about AI search.
We sell this service. Here is the evidence against most of what is sold alongside it.
The short answer
Nobody can promise you a place in an AI answer, and the published research does not support most of what is sold as GEO. We sell this service. We would still rather you knew which parts are measured and which are repeated.
The 40% figure does not mean what you were told
Almost every agency selling AI search quotes a 40% visibility lift from the original 2024 GEO paper. Olivier Martinez's critical survey of 45 GEO studies, from July 2026, spells out what that number describes: a relative gain inside a simulator where 5 documents have already been placed in the model's context.
It is not a finding that rewriting your page gets you found 40% more often. Anyone quoting it as though it were has not read past the abstract.
97% of llms.txt files have never been read
The tidiest example of a tactic sold with nothing behind it. Ahrefs checked 137,210 domains over May 2026. About 28% publish an llms.txt file, and 97% of those files received zero requests in a month. Of the 3% that were fetched, most of the traffic was SEO audit tools, not AI crawlers.
Google's Gary Illyes said Google does not support it and has no plans to. We still add the file, because it costs nothing, and we say plainly on the AI search page that no major provider has been shown to read it.
Most of the work is not on your website
This is the uncomfortable one. In studies of what AI assistants cite, content on the business's own site accounts for roughly 2% of citations. Wix Studio's AI Search Lab read 1 million citations: in professional services, third-party listicles took 80.9% of citations against 19.1% for a company's own.
So the highest-leverage work is mostly getting named somewhere else: directories, local press, a roundup, a forum thread, a video. An agency that sells you AI search and only ever touches your own pages is selling you the 2%.
The numbers move faster than the advice
Ahrefs measured how many AI Overview citations came from Google's top 10 results. In July 2025 the figure was 76%. 7 months later the same measurement gave 38%.
That is not a contradiction. It is the field moving under everyone, and it is why we date what we publish and revise it rather than leaving it up.
What we do not know
We have no data on Claude. Essentially every published study covers ChatGPT, Perplexity, Gemini and Google's AI Overviews. If somebody tells you how Claude picks its sources, ask where the number came from.
We also have no data on Albanian-language or Italian-language queries. Every study we have read is in English, on mostly American sites. For a shop in Durres that gap is not academic.
What is actually worth doing
Answer the question in the first 100 words, under a heading that asks it. Be specific: names, numbers, dates and places. Extractable facts are what gets quoted, and formatting on its own does very little.
Keep the page maintained, because freshness by last-updated date is one of the few signals that holds up. Then go and get mentioned somewhere that is not yours.
None of that is exciting and all of it is checkable, which is the difference between this and a promise.
Try it on a real business
Ask ChatGPT where to get a watch repaired in Durres, then ask it for a watch shop in Durres. We built watch.al and we would rather you ran that check than believed a screenshot.
If somebody has quoted you a number about AI search, send it over with your website and we will tell you where that number came from. Get a free audit.
Updated