How AI Search Decides Which Pages to Cite (and How to Earn a Citation)

Short answer: AI search features do not have a separate index or a secret ranking. They run searches, pull in the pages that already rank for a set of related queries, read the specific passages that answer the question, and link to the pages that supported the answer. If your page is indexed, says something clearly and backs it up, it is eligible. If it restates what ten other pages say, it is interchangeable, and interchangeable pages rarely get picked.

That is the whole game in one paragraph. The rest of this post is about what each of those steps looks like in practice and what you can realistically do about them.

What actually happens between a question and an answer

Google has been unusually open about the mechanics. Its guide to optimizing for generative AI features names two techniques behind AI Overviews and AI Mode.

The first is retrieval-augmented generation, or grounding. The model does not answer from memory. It asks Google's normal ranking systems for relevant, current pages, reads the specific information on those pages, and writes a response that is supported by them. The links you see under an AI answer are the pages that did the supporting.

The second is query fan-out. A single question gets expanded into several related searches that run at the same time. Google's own example: "how to fix a lawn that's full of weeds" might fan out into "best herbicides for lawns", "remove weeds without chemicals" and "how to prevent weeds in lawn". Each of those sub-searches retrieves its own candidate pages, which is why the links under an AI answer are often more varied than a classic results page for the original query.

ChatGPT, Perplexity and the other assistants work on the same principle when they have web access: issue searches, read, answer, cite. The search index behind them differs, the behaviour does not.

Two consequences follow. You can be cited for a question nobody has ever typed into a search box, because a fan-out query matched your page. And you can rank well for the head term and still never be cited, because the passage the model needed was not on your page or was buried in a way it could not use.

Step one is boring: be in the index and eligible for a snippet

Google's eligibility rule for its AI features is one sentence long: the page must be indexed and eligible to appear in Search with a snippet. There are no extra technical requirements, and Google says so explicitly on its AI features page.

In practice this trips up more sites than you would expect:

• A nosnippet or max-snippet:0 tag on templates. Some CMS plugins set these on category or author pages, and sometimes on everything. No snippet means no passage for the model to use.
• Robots.txt rules or a CDN bot-protection setting that blocks Googlebot from the very content you want cited. Google's guidance specifically calls out checking "any CDN or hosting infrastructure".
• Important text that only exists inside an image, a PDF or a tab that loads on click. If the sentence that answers the question is not in the HTML as text, it is not available to be quoted.
• A new Search Console setting. Since mid-2026 every property has a Search generative AI control under Settings. The default is "include". If someone on your team flipped it to exclude, your pages stop appearing in AI Overviews and AI Mode entirely, and the generative AI performance report shows nothing.

Check these first. They are the only hard gates.

Step two: write the passage the model is looking for

Retrieval works at the level of passages, not pages. When the system fans out to "remove weeds without chemicals", it is looking for a chunk of text that directly addresses that, with enough context around it to be trusted.

This is where the advice gets concrete:

• Put a direct answer near the top of each section. Open a section with the conclusion, then explain. A reader who skims gets the point, and a system that is extracting a supporting passage gets something it can quote cleanly.
• Make headings mean something. "Our approach" tells nobody anything. "How long a boiler service takes" is a question someone asks, and the paragraph under it should answer it in its first two sentences.
• One idea per section. A section that drifts across four topics is hard to quote because any extract drags in unrelated sentences.
• Use the exact words people use, once, naturally. You do not need every synonym. Google says plainly that its systems understand synonyms and that you do not have to capture every variation of a phrase. One clear statement beats five awkward ones.
• Keep facts close to their evidence. "Most domestic boilers need servicing once a year, which is the interval every major manufacturer sets as a warranty condition" is quotable. "Regular servicing is recommended" is not.

None of this is new. It is what a good editor would ask for anyway. The difference is that the payoff is now visible: pages that read like this get lifted into answers, pages that read like brochures get skipped.

Step three: say something the other candidates cannot

Here is the part most "AEO" advice skips. The retrieval step returns several candidate pages for each sub-query. If all of them say the same thing, the system can cite any of them, and the one with the strongest existing signals usually wins. You cannot structure your way out of being a commodity.

Google's guide calls this out directly. It contrasts commodity content ("7 Tips for First-Time Homebuyers", which could have been written by anyone) with non-commodity content ("Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line"), and says the unique point of view is likely to matter more than anything else in the guide.

What does non-commodity look like for an ordinary business?

• Specifics only you have: your actual prices or price ranges, your turnaround times, the three questions every customer asks on the phone, the mistake you see in half the jobs you take over.
• Positions: "we don't recommend X for Y, here's why" is memorable and citable. Hedged on-the-one-hand prose is neither.
• Original observations from doing the work: what a job looked like before and after, what the customer thought the problem was versus what it turned out to be.
• Current information: a page updated when the thing it describes changes, with the date shown, beats an evergreen page that has quietly gone stale.

If a page could be swapped for a competitor's without anyone noticing, fix that before worrying about markup.

What you can skip

Google published a mythbusting list in May 2026 and it is worth taking at face value, because it removes a lot of busywork:

• llms.txt and other AI-specific files. Google Search ignores them. Keeping one for other tools is harmless, but it does nothing for AI Overviews or AI Mode.
• "Chunking" your content into tiny pieces. Not required. Google's systems can find the relevant part of a long page.
• Rewriting pages in a special style "for AI". Also not required, for the synonym reason above.
• Chasing mentions of your brand across forums and blogs. Inauthentic mentions are treated like any other spam.
• Piling on structured data. It is not required for AI features and there is no special schema for them. Keep using it where it earns rich results, nothing more.
• Creating a separate page for every fan-out query you can think of. Google names this specifically as a violation of its scaled content abuse policy, and notes that more pages do not make a site more relevant.

If a vendor is selling you any of these as the key to AI visibility, Google's own documentation disagrees with them.

How to tell if it is working

Search Console now has a Generative AI performance report, rolled out to all properties at the end of August 2026. It shows impressions from AI Overviews and AI Mode by page, country, device and date. It does not show clicks, queries, or which of the two features produced the impression, so treat it as a visibility signal, not a traffic report.

For assistants outside Google, there is no report. The practical method is to keep a short list of the questions your customers actually ask, run them through the assistants you care about every few weeks, and note whether your page appears as a source. It is manual, but it is the ground truth, and the pattern you see after a month usually points straight at which pages are quotable and which are not.

Questions people ask

Do I need to rank number one to be cited?

No. Fan-out means the model is searching for several related things, and a page that ranks fifth for the original question can be the best match for one of the sub-queries. Being indexed, clear and specific matters more than the position for the head term.

Does AI search prefer long pages or short ones?

Neither. Google says there is no ideal page length and that its systems understand multiple topics on one page. Write as much as the question needs and no more.

Will blocking AI crawlers protect my content?

For Google, there is no separate crawler for AI features in Search; Googlebot's access is the control, and blocking it removes you from Search as well. Google-Extended is a separate token for Gemini training and grounding outside Search. For other assistants, blocking their bots removes you from their answers, which is a business decision, not an SEO one.

If you would rather not do this yourself

Writing pages that say something specific, keeping them current, and checking what the assistants actually cite is steady work rather than a one-off. It is also exactly what Ranked's SEO service does every week, including tracking how your brand shows up across ChatGPT, Gemini, Perplexity and the rest. If you want to see where you stand first, the free AI visibility checker takes about a minute.