What Are Google AI Overviews and How Does a Website Enter Them?
Published · Last updated · Written and published by Precise Search SEO, Leskovac, Serbia.
An AI Overview is a generated summary Google shows above the conventional results, with links to the sources it was assembled from. A website enters one when a page of it offers a passage answering one of the sub-questions the query was broken into — not by acquiring special markup or a particular position. An extractable passage improves the odds; it does not guarantee selection or citation.
This page explains what an AI Overview is, how it differs from AI Mode, and how pages that get in differ from pages that do not. How entry is measured, and where the limits of the report lie, is covered separately on the page about measuring results.
How does an AI Overview differ from AI Mode?
An AI Overview is a summary above the list of results on a conventional search page. AI Mode is a separate mode in which the whole page is a generated answer, with follow-up questions in the same thread. Both rest on the same web content and Search systems.
| Property | AI Overview | AI Mode |
|---|---|---|
| Where it sits | Above the conventional results | A separate mode, the whole page |
| Conventional results | Remain below | Give way to the answer |
| Follow-up questions | A new search | A continuation of the same thread |
| Typical query | Shorter, a single question | Longer, several conditions at once |
| Underlying source | The same web content and Search systems as conventional search | |
For work on a website the difference matters less than it appears. In both cases a passage is selected rather than a document, so preparation comes down to the same thing: the answer has to exist as a self-contained unit.
Where do AI Overviews draw their sources from?
From the same web content that produces conventional results. Google states that generative features are grounded in the core ranking and quality systems of Search. There is no separate index entered by a special route.
The consequence is that no separate set of techniques exists for AI. A website that is not available for crawling, has no useful content, or breaches Search guidelines will not enter an AI answer either. Everything that applies to conventional search applies here, plus one further condition: the answer must be extractable.
- Search systems — the same ranking and quality systems for every surface.
- Conventional results — a list of links.
- AI Overview — a summary above the results, with links to sources.
- AI Mode — a separate mode in which the answer is the whole page.
How is a passage found in the first place: what is multi-vector retrieval?
Classical retrieval represents a whole document as one vector and compares it with one query vector. A multi-vector approach represents a document as a set of vectors — roughly one per part of the text — so a query can match one specific part instead of an average of the whole document. This is why a passage, rather than a document, is what enters an AI answer.
The problem with the multi-vector approach is cost. Comparing a set against a set is far slower than comparing two vectors, which for a long time made it impractical at web-index scale.
What MUVERA is
MUVERA is an algorithm published by Google Research on 25 June 2025. It compresses the whole set of vectors describing a document into a single fixed-length vector, called a fixed dimensional encoding. The dot product of two such vectors approximates the true multi-vector similarity, so retrieval can run on existing fast single-vector search infrastructure.
| Approach | How a document is represented | Limitation |
|---|---|---|
| Single-vector | One vector for the whole document | Fast, but loses the difference between parts of a document |
| Multi-vector | A set of vectors, one per part of the text | More precise, but expensive to compare at scale |
| MUVERA | The set compressed into one fixed-length vector, then exact rescoring on a narrower candidate set | Two-stage: fast retrieval, then accurate re-ranking |
The procedure has two steps. The first pulls a narrow candidate set out of the full index using the compressed vectors. The second re-ranks those candidates with the true multi-vector similarity measure. The result is single-vector speed with multi-vector accuracy.
What this means for a page
If a document is compared part by part, then each part has to stand on its own. A paragraph that only makes sense alongside the one before it is a weaker candidate than a paragraph carrying the whole answer. This is the same conclusion the share-by-page-type figures point to.
- One section, one question. Not two questions in the same paragraph.
- No pronouns pointing outside the section: “this”, “that”, “as we saw above”.
- A table row understandable without the header above it and without the rest of the table.
- A section heading that names what the section is about, since it sits in the same stretch of text.
The limit of this claim: MUVERA is documented Google Research work and its description is public. Google has not stated that the algorithm is used in Search, in AI Overviews or in AI Mode. This section therefore explains how multi-vector retrieval works as a technique, not how Google Search works. The recommendations above do not depend on which algorithm is in use — they follow from the fact that a passage is selected, which is visible in the AI answers themselves.
Source: Google Research, “MUVERA: Making multi-vector retrieval as fast as single-vector search”, published 25 June 2025, with the accompanying arXiv paper. Checked 23 August 2026.
Why does the share of impressions differ by page type?
Because what gets selected is a passage carrying a concrete figure. On the taxi.co.rs project, pages whose main content is a fare table enter AI answers in 29.1% of their impressions, the site-wide average is 13.5%, and city hubs offering an overview rather than a figure stand at 6.8%.
| Page type | Main content | Share of AI impressions |
|---|---|---|
| City fare pages | Fare table | 29.1% |
| Site-wide average | All pages combined | 13.5% |
| City hubs | Overview and navigation | 6.8% |
Method. Source: the Search Generative AI performance report in Google Search Console. An AI impression is recorded when a link to the site is shown to a user in a supported generative AI feature. Property: taxi.co.rs, a project owned by the author of this page. Period: – . Aggregation: by page. Data exported . Each share is AI impressions divided by total impressions for the same page group over the same period.
On verifiability: the figures are proprietary account data. This page does not publish the underlying query set or the URL-level sample, so the results should be treated as directional rather than causal evidence.
The difference tracks what the page offers as a unit. A fare table is a unit — one row answers one question without the rest of the page. A hub offers an overview, and an overview cannot be lifted out as an answer.
What these numbers are, and what they are not: the difference between page types is measured; the cause is not. Content, layout, internal links and structured data were changed at the same time on that project, so the result is attributed to the system rather than to any single component. A control case on another project produces a markedly weaker result and is described in the leskovac.taxi study.
What is actually done on the page?
The answer to the question in the heading is placed directly below it, as a self-contained unit readable without the rest of the text. The most important figure goes into a table rather than a paragraph. Every number carries a period, a tool and a data owner.
- Phrase the section heading as a question someone actually asks.
- Write the first sentence below the heading so that it answers that heading and makes sense on its own.
- Move the most important data point into a table with
captionandscope. - State the period, the tool and the data ownership next to every number.
- Do not insert blocks between the heading and the first paragraph — they interrupt extraction of the main content.
- Use structured data to describe only what is visible on the page.
The full list of components, with a check for each, is on the page about preparing a website for AI search. Why layout alone changes which part gets extracted is set out on the page about the main content.
What must not be claimed about AI Overviews?
No structured data type guarantees a citation. No confirmed link exists between ranking position and entry into an AI answer. No official report shows the queries behind AI impressions.
| Claim | Status |
|---|---|
| An “AI Overview schema” secures a citation | False. No such type exists. |
| First position means entry into an AI answer | Unconfirmed. A passage is selected, not a document. |
| Search Console shows queries for AI impressions | False. The report gives impressions, without queries. |
| Appearing in an AI answer brings a click | Not necessarily. Clicks are not a separate metric in that report. |
| Longer text stands a better chance | It is not length but the extractability of the answer. |
The rules for drawing a responsible conclusion from any measurement are on the page about measuring results.
Common questions about AI Overviews
What is the difference between an AI Overview and AI Mode?
An AI Overview is a summary shown above the conventional results on a search results page. AI Mode is a separate mode in which the whole results page is a generated answer, with the option of asking follow-up questions. Both rest on the same web content and Search systems.
Where do AI Overviews draw their sources from?
From the same web content that produces conventional results. Google states that generative features are grounded in the core ranking and quality systems of Search, so there is no separate index entered by a special route.
Can a website be excluded from AI Overviews?
Controls exist for how content is used in generative features, but they are not without consequence. Opting out removes the possibility of appearing in an AI answer, and in some cases affects other surfaces too. The decision is made per domain, after checking exactly what is lost.
Does a high ranking position mean entering an AI Overview?
Not automatically. What is selected is a passage answering a specific sub-question, not the document with the best position. A page ranked lower that holds a clear, self-contained answer can be chosen over the page above it.
Does appearing in an AI Overview bring a click?
Not necessarily. The Search Console report for generative features gives impressions, not clicks as a separate metric. A user who has received the answer may finish the task without visiting the site.
Primary sources
This guide is a method. How it is carried out on a real website, with a price and a timeline, is shown by the semantic SEO audit.
Are your pages entering AI answers?
Send the URL and how you currently measure success.