How AI Search Actually Decides Who to Recommend (the Switch and the Crank)

TL;DR: ChatGPT does not read your website, admire your setup, and decide you are the best. It searches the web, pulls passages from other people’s pages, and repeats whoever those pages name. So the on-site “GEO” checklist most agencies sell you (schema, llms.txt, tidy markup) barely moves anything. What moves it is being present in the sources the engine pulls from: roundups, review sites, and YouTube. I will show you the evidence and what to do with it.

How does an AI engine pick who to recommend?

It runs a search and repeats what it reads. Under the hood, engines like ChatGPT, Perplexity, and Google AI Overviews use retrieval-augmented generation: a classifier decides whether to search at all, the query gets fanned out into several sub-queries, the engine retrieves passages from the top results, and the model writes an answer grounded in those passages (Ahrefs’ RAG explainer). The part that matters for you: it is citing passages it just retrieved from the open web, not judging your homepage on its merits.

That is why, when I asked four engines for the “best” software across five categories, the pages they cited were almost all third-party roundups, review sites, and YouTube, not the brands’ own sites. The engine is a repeater. Your job is to be in the sources it repeats.

Does adding schema or an llms.txt file get you cited?

No, and there is a controlled study on it. Ahrefs ran a difference-in-differences test on 1,885 pages that added JSON-LD schema against 4,000 matched controls, and adding schema moved AI citations by −4.6% on Google AI Overviews and roughly zero on ChatGPT and AI Mode. Google says the same thing in its own 2026 guidance: you do not need special markup, machine-readable files, or content chunking to show up in its AI search. Google even deprecated FAQ rich results in June 2026. The llms.txt file sits in the same inert bucket.

One honest caveat: the Ahrefs test used pages that were already heavily cited, so it does not prove schema is useless for a brand-new page trying to get discovered. But nobody should be selling schema as the thing that gets you recommended. The few on-page moves that do have evidence behind them, adding statistics, sources, and a front-loaded answer, are a separate topic, covered in what makes an AI engine quote you.

What actually correlates with getting recommended?

Off-site brand presence, by a wide margin. Across Ahrefs’ studies of tens of thousands of brands, the strongest correlates with AI visibility are YouTube mentions (~0.74) and branded web mentions (0.66 to 0.71), while backlinks manage only about 0.22. Being talked about, off your own site, correlates roughly three times stronger than the link building most budgets still chase. This lines up with what the engines actually cite: the roundups and review pages that mention you, and the YouTube videos that review your category.

What is the switch, and what is the crank?

The switch is your own site; the crank is everything said about you elsewhere. The switch decides whether an engine can read you: it needs crawler access and a page that loads. It has two positions and it has to be on, and flipping it harder does nothing. The crank decides whether you get recommended: it is your presence in the third-party sources the engine retrieves. When I audited 33 of the tools these engines recommend, about half had no valid llms.txt and many had no schema, yet they get named every time, because they are the brands the roundups and reviews talk about.

What should you do instead?

Leave the switch alone once it is on, and spend the effort on the crank.

• Get into the roundups and “best X” lists in your category. They were the biggest single citation source in my test, and many are small, reachable blogs.
• Earn unprompted brand mentions, the strongest measured correlate: Reddit, communities, podcasts, PR, other people’s content.
• Get on YouTube. It correlated highest of anything, and most small software companies ignore it as a citation channel.
• Publish your own comparison and “best for” pages with real data. Comparison pages have the highest citation-per-retrieval rate of any page type (DeltaV). We did one ourselves: an honest comparison of the AI visibility tools.

FAQ

Do I need an llms.txt file? No. About half of the AI-recommended tools I audited do not have a valid one, and Google says it is not required.

Is on-site SEO now useless? No. You still need the switch on: crawlable, loads, exists. Just do not expect on-site tweaks past that to move recommendations.

Which engine matters most? They overlap on the winners but pull from different sources and reshuffle run to run. Optimize for the off-site presence they all draw on, not one engine.

How fast does this change? Fast. ChatGPT’s use of Reddit and Wikipedia shifted sharply in late 2025. Treat any specific number here as a 2026 snapshot.

Sources and method

Figures come from Ahrefs’ schema difference-in-differences study and brand-correlation studies, Google’s 2026 AI-optimization guidance, and DeltaV’s citation study, plus my own test running ChatGPT, Gemini, Claude, and Perplexity across five categories and auditing 33 of the tools they recommend. Correlational findings are correlational, most sources are 2026 snapshots of fast-moving systems, and I have flagged the one causal study’s limits above. If your category is one where AI referrals are starting to matter, which sources is it pulling from when it names your competitors instead of you?