2026-09-03 · 4 min read
Google AI Overviews: how to appear in them, and how to know whether you do
AI Overviews answer above the results and cut the clicks. What they read, what gets you into them, and why checking by hand is not enough.
Since Google started displaying a written answer above its results, the question has changed in kind. It is no longer about being in the ten links: it is about being in the paragraph that replaces them.
This article answers three concrete questions: what AI Overviews are, what determines the brands that appear in them, and how to check your situation without kidding yourself.
AI Overviews, SGE: one subject, two names
Google launched the feature under the name SGE — Search Generative Experience — in testing. It was renamed AI Overviews when it opened to the public. If you are searching for "google sge", you are searching for the same thing: the written answer block at the top of the results page, with a few source links alongside.
So there are not two optimisations to run. The term SGE belongs to the past; the object is identical.
What changes for you, in one sentence
A classic result gave you a click. An AI Overview gives you a mention, and sometimes no click at all, since the answer is already there.
That is a change of unit, and it is what makes most existing dashboards blind: they count positions and clicks. A mention inside a written text is neither.
What the AI Overview reads before writing
The mechanism is that of every generative engine: reformulate the question into several queries, read a small number of pages, synthesise, name two or three.
Our measurements cover ChatGPT, Perplexity and Claude, not AI Overviews — we say so because the distinction matters, and because articles claiming otherwise are usually reconstructed from three screenshots. But the selection mechanism is the same, and the finding it produces carries over: it is rarely the brands' own sites that get read.
Across 88 measured answers about campsites in Normandy, the most-cited name in the entire market is not a campsite: it is the regional tourist board, present in 18% of answers and across 9 of the 10 questions asked. On one question, the five sources consulted were five tourist boards — not a single property read directly. The full study, with sources.
Google's own particularity is real but limited: the AI Overview draws on Google's index, where Perplexity or ChatGPT run their own queries. In practice that means a site invisible in classic SEO is also invisible to the AI Overview — the old discipline remains the entry condition, it is simply no longer the exit condition.
The three levers, in order
1. Be named on the pages that are themselves well ranked. If the AI Overview reads the top three results for a query and none of them names you, you will not appear — however well optimised you are. Sector comparisons, directories, specialist press, forums, local bodies: the question is not "how do I climb", it is "who among those already at the top talks about me".
2. Answer the question, literally. AI Overviews trigger mostly on questions. A page that puts the question in the heading and answers in the next sentence is quotable as is. A page that buries the answer in the fourth paragraph is not.
3. Write attributable facts. A figure, a date, an area, a duration, a compatibility. A model reproduces a promise badly and a data point well. "Delivery within 24 hours in mainland France" gets cited; "ultra-fast delivery" does not.
Why checking by hand is not enough
You type your query, you look at the AI Overview, you do not see yourself. You have learned nothing.
Generative blocks do not trigger systematically; they vary with the account, the location, the phrasing, and they change from one run to the next. It is the same problem as on the assistants: the same question asked three times gives three answers, sometimes naming different brands. Why one measurement proves nothing.
A check that stands up needs several questions, several passes, and questions that do not contain your name — otherwise the answer talks about you because you asked it to.
What we measure, and what we do not
Our GEO audit covers ChatGPT, Perplexity and Claude: 40 questions from your market, three passes each, 360 measurements. It does not cover AI Overviews today, because there is no stable, honest access to those blocks — scraping them would produce an unverifiable figure, and we would rather not sell that.
What the audit does give you, and what holds for AI Overviews too: the sources actually read in your market. It is the same list of pages to target in both cases, and it is the part of the work that does not depend on the engine.
For the underlying subject, see the GEO guide and AI search optimization in practice.