Google's AI Overviews: how to appear as a source

Above many Google results now sits a generated answer box: the AI Overview. It answers the question right on the results page, with a few source links beside it, and pushes the classic results down, often by a full screen height. For website owners this is the most consequential change to Google search in years, and it triggers two reactions: panic about the clicks, or a shrug. Both are wrong.

What is true: the AI Overview costs clicks, and at the same time it is a new surface you can stand on. This piece explains when the overview appears, how Google picks its sources, what really happens to the clicks, and which work raises the odds of being cited. Plus the two defensive moves that do not work, although they are recommended everywhere.

When the overview appears and when it does not

The AI Overview does not show up for every search. Google displays it mostly where a question can be summarized from several sources: explainer questions, comparisons, "how do I" tasks, health and advice topics. It rarely appears for searches with clear purchase intent, for navigational searches for a specific brand, and for local searches, where map and profiles still dominate.

That distributes the impact very unevenly. A website living on advice traffic sees the overview above most of its important queries. A shop or a local business notices far less. Before you invest time: type your ten most important queries by hand and count how many carry an overview. That is your actual exposure, and your actual opportunity.

How the overview is built

Google does not answer from the model's gut; it builds the overview from search results. The process resembles what ChatGPT with web search does, only with Google's own index: the query is expanded in the background into several sub-questions (Google calls this query fan-out), the top results for each are read, and Gemini writes the answer from the most usable passages. The pages whose statements were used appear as source links next to or below the text.

From this follow the two conditions almost everything hangs on. First: you must rank high enough for at least one of the sub-questions to be read at all, because the overview overwhelmingly cites pages that already place well. Second: your page must contain a section that answers one of those sub-questions completely, because passages get picked, not domains.

This also explains why overviews regularly cite pages sitting at position eight or twelve in the classic list: their passage was the best answer to a sub-question the searcher never typed.

Which data sources feed an overview

"Built from search results" is the short version. In practice an overview assembles itself from several pools, and it pays to tell them apart, because what you can do about each one differs sharply.

The web index carries the bulk of it: the top results for each sub-question, fetched and cut into passages. Nearly all of your leverage sits here, and the section on getting cited is entirely about this pool.

The Knowledge Graph supplies the entity facts: that a company exists, what it is called, where it sits, since when, what it belongs to. These statements often appear in an overview without a source link beside them, because they are treated as established rather than quoted. You feed this pool from three directions at once: a maintained Google Business Profile, entries on the reference platforms Google trusts, and an Organization or LocalBusiness block on your own site whose sameAs list points at exactly those profiles.

The Google Business Profile answers the local half directly: address, hours, phone, reviews. For a local business it is the most valuable data source of the lot and the one most likely to be stale. What belongs in it is covered in getting found locally.

The Merchant Center feed answers price and availability for product questions. Overviews on shopping queries take those figures from the feed, not from your page, which is why a price can be current in your HTML and wrong in the overview at the same time.

Datasets and statistical sources answer everything that wants a figure. When a question asks how much, how many or how often, an overview prefers statements that carry the number together with its unit, its period and its origin, and it prefers whoever produced the figure over whoever repeated it. Google keeps a separate index for this, Dataset Search, fed by Dataset markup. Most business websites have nothing to do here. Anyone who publishes figures of their own, a price index, a yearly evaluation, a customer survey, has the cheapest route available to becoming the source of a number instead of one more site quoting it.

If that describes you, the markup is a single block, and it belongs on the page that shows the figures, next to the visible table:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Dataset",
  "name": "Fitted kitchen prices in the South West 2026",
  "description": "Final prices of 240 fitted kitchens installed between January and December 2025, broken down by kitchen size and appliance class.",
  "url": "https://harperkitchens.com/research/kitchen-prices-2026",
  "creator": { "@type": "Organization", "name": "Harper Kitchens" },
  "temporalCoverage": "2025-01/2025-12",
  "spatialCoverage": "South West England",
  "datePublished": "2026-02-10",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "distribution": {
    "@type": "DataDownload",
    "encodingFormat": "text/csv",
    "contentUrl": "https://harperkitchens.com/research/kitchen-prices-2026.csv"
  }
}
</script>

The two fields that do the work are temporalCoverage and creator: a number without a period and an origin is not a source, it is a claim, and answer systems treat it as one. The rest of the block is bookkeeping.

The model's own knowledge writes the connective tissue between the quoted parts. It cites nothing, it cannot be addressed, and it is the reason an overview occasionally states something confidently that appears in none of the linked sources.

What really happens to the clicks

The honest number first: when an AI Overview sits above the results, considerably fewer people click the entries below. Studies such as Pew Research Center's from 2025 measure roughly a halving of the click-through rate. If you follow your impressions and clicks in Search Console, you see the pattern directly: impressions stable or rising, click-through rate falling. That is not a ranking loss; that is the overview wedging itself between you and the click.

But the equation has a second half. The clicks that still happen go disproportionately to the cited sources. Whoever stands in the overview as a source gets clicked by people who read the summary and want more. Fewer visitors, but noticeably warmer ones. And for searches where the overview delivers the complete answer, say opening hours, a definition or a date, the click was worth little before, too.

The strategic conclusion: content whose only value is a quick factual answer loses. Content that still offers a reason to click after the answer wins in relative terms: tools, calculators, first-hand experience, concrete instructions with details no summary can hold.

How to get into the overview

There is no switch and no registration. There is work that raises the probability, and it overlaps heavily with what makes pages good anyway.

Rank for the sub-questions, not just the head keyword. If your page covers the topic in its breadth, meaning process, costs, duration and typical mistakes, it is a candidate for more of the fan-out's sub-questions. A page serving only the head phrase holds exactly one ticket in the draw.

Build paragraphs that can be cut out. Every important sub-question deserves a section with a heading that names the question and a first sentence that answers it. Checkable details such as numbers, units and dates raise the pickup chance, because the model prefers statements it can attribute to a source.

Keep the facts machine-readable. Structured data makes date, author, price and question-answer pairs unambiguous. Not a free pass, but it removes ambiguities that would otherwise knock a passage out of the running.

Show freshness. For many topics, overviews visibly prefer fresh sources. A visible, maintained update date counts double here, provided actually updated content sits behind it.

Be the answer elsewhere too. Google cross-checks statements across sources. Mentions in directories, trade articles and forums back your claims; contradictory stale data across the web weakens them.

The markup that carries each kind of answer

Structured data does not get a page into an overview; ranking for the sub-questions does that. What markup settles is the question the model has to answer before it can quote you: whose number is this, and when was it true. A passage that cannot be attributed cleanly drops out in favour of one that can. Which type does that job depends on what your page answers.

• Explainer and comparison pages: Article with author and dateModified. The modification date is the underrated field, because overviews visibly prefer the fresher source on time-sensitive topics.
• Facts about your business: Organization, or LocalBusiness with address and opening hours, once, site-wide. The sameAs list is what ties your directory and profile entries back to one identity.
• Prices and availability: Product with an Offer. For shops the feed in Merchant Center outranks this, but the markup is what a system without feed access reads.
• Question-and-answer sections: FAQPage, with the caveat that Google has barely shown the expandable FAQ results since 2023. It is still read, and it still removes the ambiguity of which answer belongs to which question.
• Figures of your own: Dataset, as above.

The blocks, ready to paste, are in structured data for AI answers. Two rules outrank the choice of type. Markup may only state what stands in the visible text of the page; an annotated price that appears nowhere on screen is not optimization but a false statement, and Google treats it as one. And markup is not a ranking signal, so it cannot rescue a page that would not have been read anyway.

What does not work

Two widespread recommendations lead nowhere. First: blocking Google-Extended in robots.txt to disappear from AI Overviews. That token exclusively governs training of the Gemini models. The overviews are fed from the normal search index and remain untouched. The differences between crawlers and control tokens are sorted out in Controlling AI crawlers.

Second: cutting yourself out of the overview with nosnippet or max-snippet. That works technically, since Google may then no longer quote your content in overviews. But it simultaneously trims your normal snippets in the classic results. You vanish from the overview and become less attractive below it. For the vast majority of websites this trade is a loss; it makes sense at most for publishers who fundamentally do not want their content summarized.

How to measure what the overview does to you

Here lies the most uncomfortable property of AI Overviews: they are nearly invisible in the data. Search Console does not report impressions and clicks from overviews separately; they hide inside the normal numbers. A click from an overview arrives in GA4 as an ordinary organic Google visit, indistinguishable from a click on a blue link. Why that is, and what can be measured anyway, is covered in Measuring AI traffic in GA4.

What remains is circumstantial work with Search Console: filter to your important queries and watch for the pattern "impressions stable, click-through rate falling". That is the signature of an overview above your results. The pattern reads more reliably once Search Console sits inside GA4, because then the falling click-through rate can be held against what the remaining sessions actually do; how to link the two accounts is in connecting GA4 and Search Console. And check the queries by hand: is there an overview, and are you cited? A monthly sample across your ten most important queries says more than any curve.

Frequently asked questions

How do I get into a Google AI Overview?

By ranking well for the topic's sub-questions and offering sections that each answer one of them completely, with a heading that names the question, a direct first sentence and checkable details. There is no registration and no switch; the overview picks its sources from the top search results.

Can I remove my website from AI Overviews?

Only via snippet restrictions such as nosnippet or max-snippet, which also trim your normal search snippets, or by leaving the index entirely. Google offers no opt-out for overviews alone, and Google-Extended affects Gemini training only.

Do all websites lose clicks to AI Overviews?

No, the losses are distributed very unevenly. Advice and definition content is hit hard; searches with purchase intent and local searches rarely show overviews. Cited sources also receive a disproportionate share of the remaining clicks, from visitors who arrive already informed.

Does Search Console show whether I appear in AI Overviews?

Not directly: impressions and clicks from overviews sit unlabelled inside the normal performance data. The typical clue is stable impressions with a falling click-through rate on individual queries. Whether you are cited, only a manual sample reveals.

Do AI Overviews appear outside the US?

Yes. They have been rolling out across Europe since 2025, Germany included, in the local language and across a growing share of query types. Your own exposure is best checked with a manual sample of your most important queries, because the share of searches carrying an overview shifts constantly.

Where do AI Overviews get their data from?

Mostly from Google's normal web index: for every sub-question of the fan-out the top results are read and cut into passages. Around that sit the Knowledge Graph for entity facts, the Google Business Profile for local details, the Merchant Center feed for prices, and a separate dataset index for figures. The model's own training knowledge writes the connecting sentences but cites nothing.

Does structured data get me into AI Overviews?

No. Markup does not make a page a candidate; ranking for the topic's sub-questions does. What it does is make your facts unambiguous, so a passage does not drop out because the model cannot tell whose price, date or opening hours it is looking at.

Is Dataset markup worth it?

Only if you publish figures of your own, say a price index, a survey or a yearly evaluation. Then Dataset markup with a period and a named originator makes you the attributable source of the number rather than one of the sites quoting it. Anyone who only cites other people's figures gains nothing from it.