Google AI Overviews, and how a page ends up inside one

A Google AI Overview is a generated summary that appears above the ordinary search results, written by a Gemini model from pages Google has already indexed, with links to the sources it drew on. It is not one page quoted — it is several pages synthesised.

That distinction decides almost everything else about how you work with them. You are not trying to win a position any more; you are trying to be one of the handful of sources a model reaches for when it composes an answer. This guide covers what is actually known about how that selection works, what you can influence, and the one thing about AI Overviews that nobody — including us — can measure honestly.

What an AI Overview is

When Google judges that a query can be answered rather than merely matched, it composes a short passage at the top of the results page. The passage is generated, not retrieved: the model is given search results as grounding and writes prose from them. Alongside it sits a set of source links — sometimes inline, sometimes in a panel to the right, sometimes behind an expander.

Three properties follow from it being generated, and each one breaks an assumption carried over from ordinary SEO:

• It is composed fresh. The text is not stored and served; it is produced when asked. The same query can yield different wording and a different set of sources minutes apart.
• It cites several pages. A featured snippet has one winner. An Overview has a handful, and being one of them is a different competition from being first.
• It does not appear everywhere. Coverage varies by query type, language and country, and Google adjusts when it triggers. A query that shows one today may not tomorrow.

Not a featured snippet

The two get conflated constantly, including by tools that should know better. They are different mechanisms with different implications for the work.

The second column is harder to work with, and pretending otherwise is where most advice on this topic goes wrong. You cannot inspect a generated answer and recover the reason a page was picked. What you can do is track whether you are being picked, across enough samples that the pattern is real.

Where the sources come from

AI Overviews are grounded in Google's index — the same corpus behind ordinary results. In practice the cited sources overlap heavily with pages that rank well for the query, which is why the fundamentals still matter. But the overlap is partial, and the exceptions are the interesting part:

• Pages below the fold get cited. Sources are not restricted to the top three. A page ranking seventh can be quoted while the second is not.
• The query is decomposed. A multi-part question is answered from several sub-searches, so a page that answers one clause well can be cited even if it is unremarkable for the query as typed.
• Aggregators and communities appear often. Pages that compare options or collect opinions are useful grounding for a model writing a balanced summary, which is why a roundup you are absent from can outrank your own product page as a route into the answer.

That last point is the one most worth acting on, and it is not a page you own. If the comparison article an engine reads every time names four competitors and not you, the fastest route into the answer is being added to that article — not publishing a fifth page on your own domain.

What correlates with being cited

Everything in this section is correlation observed across many queries, not a mechanism Google has documented. Treat it as where to spend effort first, not as a formula.

Answering the question outright, early

Pages that state the answer in a self-contained passage near the top are easier to ground a generated sentence in than pages that build to it. This is the single most consistent pattern, and it is also just good writing.

Question-shaped structure

Headings phrased as the questions people actually ask give a model a clean unit to extract. A page organised as What it costs / How long it takes / What it does not do is more quotable than the same content under Overview / Details / More.

Being specific enough to be worth quoting

Numbers, dates, named constraints and stated limits survive summarisation. Adjectives do not. “Fast, reliable and affordable” carries nothing a model can attribute to you; “returns results in about ten seconds, capped at 500 pages” does.

Reachability by the crawlers that matter

Google-Extended, GPTBot and their equivalents are separate user-agents from Googlebot, and a robots.txt written years ago can block them without anyone noticing. We track 13 of them in the free crawler checker — it is a two-minute check that occasionally finds a site has been invisible to AI crawlers for months.

What you can and cannot control

The right-hand column is not a reason to disengage; it is a reason to be sceptical of anyone selling control over it. There is no submission process, no markup that admits you, and no agency relationship with Google that changes the outcome.

The measurement problem

This is the part most pages on this subject skip, and it is the part that decides whether any number you are shown means anything.

There is no public API for AI Overviews. Google does not expose an endpoint that says whether an Overview appeared for a query, which sources it cited, or how much traffic it sent. Everything any tool reports about them — including ours — is derived from sampling search results and reading what was there.

Search Console does not break them out. Clicks and impressions associated with AI Overviews are folded into the ordinary Search performance report rather than separated as their own appearance type. So you cannot open your own data and compare AI Overview traffic against everything else — the rows are already mixed together before you see them.

Analytics cannot separate them either. A visit that started in an AI Overview arrives as organic search from Google, the same as any other. There is no parameter that marks it. Any dashboard showing you an “AI Overview traffic” line is either modelling it from assumptions or labelling ordinary organic traffic as something it cannot verify.

We say this knowing it is an awkward thing for a company selling AI-visibility tracking to lead with. It is also the reason to trust the numbers we do show: we would rather state the boundary than draw a graph across it.

What you can track instead

One thing about AI Overviews is observable, and it happens to be the thing that matters most: whether you appear in one at all. That can be measured by running the query and reading the result — which is exactly what a rank tracker has always done, pointed at a different part of the page.

1. Pick the questions, not the keywords. AI answers are composed for questions. “Best CRM for a ten-person agency” is a query worth tracking; “CRM software” tells you nothing about whether you would be recommended.
2. Sample repeatedly. Because the answer regenerates, a single check is noise. The useful unit is a rate across weeks: cited in eleven of the last twelve runs means something; cited once does not.
3. Record who was named instead. When you are absent, the answer names somebody. That list is the most actionable output of the whole exercise, because it tells you which comparison pages and communities the engine is actually reading.
4. Separate the durable from the volatile. Some brands are named almost every time; others rotate in and out. Treating a lucky appearance as a win — or a single absence as a loss — leads to chasing noise.

That last distinction is the one we built the product around. Being named 80% of the time and being named once are different facts about a brand, and a tool that reports them the same way is hiding the only thing worth knowing. You can see how your own brand splits on that line with the free AI visibility checker, or run a single page through the 31-check AI readability audit to see what an engine can and cannot extract from it.

Questions people ask