Google AI Overviews and AI Mode guide

The short answer

Google AI Overviews and AI Mode are two Search surfaces, not one ranking position. AI Overviews appear on some ordinary results pages when Google judges that a generated summary adds value. AI Mode is a conversational search experience for exploration, reasoning, and comparison. Both can issue multiple related searches through query fan-out, then assemble a response with supporting web links. The same question can therefore produce different wording and different sources on the two surfaces.¹

There is no special AI Overview schema, llms.txt submission, or separate Google AI crawler to unlock. A page must be indexed, eligible to appear in Search with a snippet, and included in Google's Search generative AI features. After eligibility, selection is competitive and query-specific. An allowed crawl or valid schema can establish Eligible evidence; it cannot establish that a page was selected, mentioned, or cited.¹

How Google finds and sources supporting pages

Google says AI Overviews and AI Mode may use query fan-out: the system issues related searches across subtopics and data sources while building a response. That helps explain why a page can support one narrow part of an answer without ranking first for the wording a user typed, and why the visible source set can be wider than the first page of classic results. It does not reveal the generated subqueries, a fixed source quota, or a recipe for selection.¹

Work backwards from the information need, not a presumed prompt string. A comparison question may require definition, criteria, prices, constraints, alternatives, and current evidence. Put the claims a reader needs into clear textual content; connect the relevant pages with descriptive internal links; keep dates, authorship, methodology, images, video, Merchant Center data, and Business Profile information accurate where they apply. Structured data should match the visible facts, not extend them.

Do not publish a separate thin page for every imagined fan-out query. Google's current guidance warns that generating query-variation pages primarily to manipulate rankings or generative responses can violate its scaled-content-abuse policy. A smaller body of non-commodity content—original observations, direct experience, useful data, and claims another source can check—gives both readers and retrieval systems more reason to choose the page.¹

Publisher controls: inclusion, crawling, previews, and model use

Google now exposes a Search generative AI control in Search Console for a subset of verified properties. Its default is Include. Exclude prevents the property's links and content from appearing in the covered Search generative AI features or grounding their responses, while leaving other parts of Search unaffected. Parent and child properties can inherit or override the setting. Google says a changed exclusion generally takes a few days to propagate, with content usually excluded within one or two days after the control goes live.²

Treat that property-level choice separately from page-level preview controls. `noindex` removes a page from Google Search, not only AI features. `nosnippet` prevents a text snippet and prevents the content being used as a direct input for AI Overviews and AI Mode. `max-snippet` limits how much text may be used, while `data-nosnippet` excludes selected visible sections. Google must be allowed to crawl the page to discover these directives; a robots.txt disallow can hide the very meta rule you expect it to obey.

Google-Extended is another separate decision. It is a robots.txt product token with no HTTP user agent of its own. Google documents it for future Gemini-model training and for grounding in Gemini apps and Vertex AI; it does not affect inclusion or ranking in Google Search. It is therefore not the switch for AI Overviews or AI Mode. Search eligibility is governed through Googlebot, indexing and the Search generative AI control.^2,5

<!-- Remove this page from Google Search entirely -->
<meta name="robots" content="noindex">

<!-- Or keep the page indexed but prevent a text snippet and direct AI input -->
<meta name="robots" content="nosnippet">

<!-- Or permit a bounded text preview -->
<meta name="robots" content="max-snippet:160">

<!-- Exclude only a visible passage from snippets -->
<section data-nosnippet>Subscriber-only analysis…</section>

What Search Console measures—and what it withholds

Google's dedicated generative AI performance report covers AI Overviews and AI Mode and is still rolling out to a subset of properties. It reports impressions grouped by Pages, Countries, Devices, and Dates. An impression means a link from the property was shown under Google's visibility rules on a covered feature. The report has no query dimension, no answer transcript, no citation context, and no dedicated click or conversion measure.³

Read a rise as increased covered exposure for the property or page. It does not prove a citation, recommendation, positive description, visit, or sale. Property totals and page rows use different aggregation, the newest values can be preliminary, and exported `~` or `-` values become zeros. Keep the raw export, date range, time zone, property, filters, and aggregation with any chart so later readers can reconstruct what changed.

For query-level evidence, run a controlled observation: record the exact query, surface, locale, signed-in state, time, answer, visible source URLs, and whether the target domain appeared. Repeat it on a schedule under comparable conditions. That establishes what one surface returned at one time. It does not establish population-wide exposure or causality. Use web analytics separately for attributable visits and conversions.

First-party observation: one query, two Google surfaces

On 11 August 2026 we ran the category query in a signed-out English Google session localized to the United Kingdom. The ordinary results page produced an AI Overview that defined answer engine optimization and exposed visible source links to HubSpot, Reddit, Forbes, and YouTube.^6,8,9,10,11

We then ran the same query in the AI Mode tab. It generated a longer answer and showed BigRock, Frase, and Saffron Edge as its visible related links. Robot Visible was not among the visible sources on either surface.^7,12,13,14

The supported finding is narrow but useful: at that time, the two Google surfaces sourced the same query differently, and our domain was absent from the sources we could inspect. This is one observation, not a trend. It does not say the result is stable, that every user saw it, that those were the only sources used internally, or that Robot Visible is absent for other questions, countries, devices, or signed-in contexts.

To reproduce it, open the linked signed-out query with `hl=en` and `gl=gb`, record whether an AI Overview appears, then open the AI Mode version and capture its visible related links. Preserve the time and state even if your answer differs. A changed result is not a failed reproduction; it is a later observation that belongs beside this one.

A measurement plan that preserves the evidence

Start with an Eligible record: the exact URL returns successfully, Googlebot is allowed through robots.txt and the CDN, the page is indexable and eligible for a snippet, the canonical is correct, important claims are present in server-readable text, and the Search generative AI control is set to Include for the applicable property. Record unknown access or unavailable controls as unknown, not as a pass or failure.

Then build Competitive evidence around a small question set. For each question, list the claim types the answer needs, inspect the pages Google currently exposes, and identify what they provide that yours does not: primary experience, current data, a useful comparison, transparent methodology, stronger corroboration, clearer product facts, or a more directly supported passage. Ship the smallest change that closes a real evidence gap rather than rewriting everything around guessed keywords.

Finally collect Observed evidence. Save each answer with the exact query, provider, surface, source URL, locale, state, and timestamp. Compare like with like after the page is recrawled. Search Console impressions, manually observed sources, and analytics visits remain separate rows. Movement after a release is a correlation until the design isolates the change; a before-and-after chart alone does not establish causality.

Run a free Googlebot and page-readiness check

The practical publishing checklist

• Verify the correct Search Console property and its inherited Search generative AI control before diagnosing content.
• Keep Googlebot allowed at robots.txt, CDN, firewall, and origin layers; verify claimed bot traffic rather than trusting the user-agent string alone.
• Confirm the canonical URL is indexed and eligible for a snippet; inspect noindex, nosnippet, max-snippet, data-nosnippet, and X-Robots-Tag together.
• State the answer and its evidence in visible text. Add original experience, data, examples, methodology, dates, authorship, and source links where they help a reader.
• Use structured data only when it matches the visible page. Keep Merchant Center and Business Profile facts current when those data sources apply.
• Choose a fixed, representative query set. Record AI Overview absence as 'not triggered', not as a lost citation.
• Run AI Overview and AI Mode as separate surfaces, preserving locale, signed-in state, answer, visible source URLs, and timestamp.
• Read Search Console generative AI impressions as aggregate exposure; keep answer observations and analytics outcomes separate.