How Google AI Overviews and AI Mode choose their sources
Google AI Overviews and AI Mode draw their sources from the same index as ordinary search results, so a page must be indexed and eligible to show a snippet before it can be cited at all. Google documents no extra technical requirement beyond that — no special file, no dedicated markup — but eligibility is not selection, and the page cited is frequently not the page ranking first.
The AI answer sits on top of search, not beside it
This is the single fact that reorganises everything else. Google's AI answers are not a separate product with a separate index you need to get into. They are generated over the same index that produces the ten blue links underneath, using the same crawling and the same eligibility rules.
Google's own documentation on AI features is unusually direct about this: to be eligible as a supporting link, a page needs to be indexed and eligible to appear in Search with a snippet, and there are no additional technical requirements. It also says explicitly that no new machine-readable files, AI text files or special schema markup are required.
That makes Google the assistant with the most overlap with work you are probably already doing — and the one where the most money gets wasted, because the market keeps selling separate "AI optimisation" products for a surface that runs on your existing indexing.
Eligibility is not selection
Being eligible and being chosen are different problems, and conflating them is the most common mistake we see. Thousands of pages are eligible for any given question. A handful get linked.
Google describes these features as using query fan-out: rather than answering the one question typed, the system issues multiple related searches across sub-topics and data sources, then assembles a response from what comes back. Google says this is what lets it surface a wider and more varied set of links than a classic results page.
Think about what that does to your targeting. A dental practice in Leeds optimising hard for "invisalign Leeds" is competing for one query. The fan-out behind a real user question — "is invisalign worth it or should I get fixed braces" — pulls on cost, treatment time, suitability, aftercare and local availability. The cited pages are the ones that answered those pieces. A page that answers a sub-question well can be cited while the page ranking first for the headline term is not, which is the mechanical reason you can rank on Google and still not be named.
Being cited is not the same as ranking first
Two further asymmetries are worth internalising.
First, position in the AI answer does not mirror position in the results. The links shown alongside a generated answer are drawn from across the fan-out, not from the top of a single ranking, so your organic position is a weak predictor of whether you appear.
Second, AI Overviews do not always appear. Google says they are shown only when its systems judge that they add something beyond the ordinary results, so coverage varies by query type and by market. Google's AI features have also rolled out at different times in different countries, with different languages and different availability — so a competitor in another market reporting a dramatically different experience may simply be describing a different rollout. If you operate across borders, read how AI answers differ by country before drawing conclusions from a single market's results.
How the two features differ in practice
AI Overviews and AI Mode are not the same surface, and the difference is worth holding even though the details of each will keep moving.
An AI Overview is a generated summary placed above a conventional results page that still exists beneath it. The user came for search and got a summary on top. AI Mode is a conversational surface in its own right, where the answer is the page and the follow-up question is expected. Google has said the two may use different models and techniques, so the links they show can differ for the same question.
The consequence for a business is about depth of intent. An Overview tends to resolve a first, broad question. A conversational exchange narrows: someone three questions into a thread about hiring a commercial lawyer in Austin is a far better prospect than someone who typed the category once. Content that answers the narrower, later question is competing for better attention, even though it will never show impressive keyword volume.
What this does to clicks
An answer that resolves the question in place removes the reason to click. That is the honest summary, and it applies unevenly: informational queries lose the most, and queries where the user needs to transact, book or compare specifics still produce visits.
We will not put a percentage on the decline, because the published figures vary wildly by methodology, by sector and by market, and a confident number here would be a guess dressed up as data. What is reliably observable in most accounts is a widening gap between impressions and clicks — the same visibility converting into fewer sessions. The pattern and its implications are covered in the zero-click problem.
The strategic consequence is not that clicks stopped mattering. It is that being named inside the answer now has standalone value, because for a share of users the answer is the whole interaction. A citation that produces no visit still put your name in front of someone who was asking a buying question.
Google-Extended, and the controls that actually apply
Google-Extended is widely misunderstood, including by people writing robots.txt advice. Google documents it as a control for limiting AI training and grounding in some of its other systems. Its guidance for AI features in Search points site owners elsewhere: to ordinary Googlebot crawling directives, and to the snippet controls.
So the levers that genuinely affect what Google can show from your pages in an AI answer are the familiar ones:
- Googlebot directives in robots.txt govern crawling for Search, and AI features are built into Search.
- noindex removes the page from the index entirely, and therefore from eligibility.
- nosnippet, max-snippet and data-nosnippet limit what can be shown from the page, including in previews. These are blunt: restricting snippets to protect content also restricts what can be surfaced about you.
Blocking Google-Extended while leaving Googlebot alone does not opt you out of AI Overviews. Blocking Googlebot opts you out of Google altogether. Neither of those is usually what the person asking wanted, which is why getting your robots.txt right for AI crawlers is worth half an hour of care rather than a copied snippet.
What practically helps
Because the eligibility bar is ordinary search eligibility, the work divides cleanly into what you already do and what is genuinely new.
- Keep the indexing fundamentals solid. Crawlable, fast, indexed, snippet-eligible. Nothing here replaces that, and a page that is not indexed cannot be cited.
- Write for sub-questions, not just head terms. Fan-out rewards pages that answer a specific thing completely. One honest page on what a treatment actually costs in your area will be pulled more often than three pages tuned to a keyword.
- Make statements extractable. A generated answer is assembled from sentences it can safely restate. Vague positioning copy gives it nothing to use.
- Qualify your location properly. Mansfield exists in England, Texas and Ohio; Hamilton, Perth, Richmond and Birmingham all exist in several countries. An unqualified town name invites an answer about the wrong continent.
- Do not buy a separate technical format on the promise of AI inclusion. Google's own documentation says no new file or markup type is needed for these features.
One thing that is new rather than borrowed from SEO: the unit of work shifts from the page to the claim. A results page rewards a page that is comprehensively about a topic. A generated answer rewards a sentence that can be restated without the surrounding context and still be true. Those are not the same target, and the second one is why a thorough, well-ranking service page can lose to a short, specific paragraph on a competitor's blog.
Measuring it without fooling yourself
Google began adding generative AI performance reporting to Search Console during 2026. As of late 2026 that reporting is impression-based rather than click-based, and it has not reached every property, so it answers "was I shown" far better than "did it do anything".
A workable measurement stack: Search Console for whatever AI impression data your property has, your analytics for referred sessions, and a repeated set of real buying questions asked on a schedule for the part neither of those can see. None of the three is complete. Together they are enough to tell whether you are moving.
One caution to carry through all of it. These systems are probabilistic and the rollout is still changing, so a single check — good or bad — is close to meaningless. What matters is the same questions, asked repeatedly, over enough weeks that the noise averages out. A free visibility check is a reasonable way to set the first marker before you start changing anything.
Common questions
Does ranking first mean I will be cited in the AI answer?
No. Ranking makes you eligible, not chosen. Google describes these features as issuing multiple related searches around the original question, so the cited pages often answer a sub-question rather than the headline query. A page at position seven that answers the specific sub-question can be cited while position one is not.
What does Google-Extended actually control?
Google documents Google-Extended as a control for limiting AI training and grounding in some of its other products. For AI features inside Search, Google points site owners to ordinary Googlebot crawling directives and the snippet controls instead, so Google-Extended is not the lever for AI Overviews or AI Mode.
Do I need llms.txt or special schema to appear?
Google states on its own documentation for AI features that no new machine-readable files, AI text files or special schema.org markup are needed, and that standard search fundamentals still apply. Structured data remains useful for other reasons, but it is not an entry requirement for being cited here.
Can I see how often I appear in Google AI answers?
Partly, and only recently. Google began adding generative AI performance reporting to Search Console during 2026. As of late 2026 that reporting is impression-based rather than click-based and has not reached every property, so treat it as one input alongside your own repeated checks.
See whether the assistants name you
Three real customer questions for your trade and town, no signup, about a minute.
Run my free check