Google Search Central AI Features and AI Overviews Optimization Documentation

Google Search Central's AI features and AI Overviews optimization documentation is two pages plus a Search Console announcement. Teams ask for a "documentation brief" when someone on the account pasted a Twitter thread over those URLs and now wants the thread treated as the spec. We read the primary pages. We do not add ranking factors Google did not write, because an invented factor billed into a retainer is worse than no factor at all.

The Twitter-thread-over-the-URLs problem is the one that produces the documentation brief request, because a thread is shorter than the documentation and easier to paste into a Slack channel. The thread is not the spec. The documentation is the spec, and the documentation is the thing Google wrote. A team that treats the thread as the spec builds a program on a claim Google did not make, and the claim is the part that gets caught in a review. The discipline is to read the primary pages and quote them, which is slower than pasting a thread and produces a program that survives a question about where the claim came from.

This is the documentation index. The practical reading with measurement notes is AI features search content guidance. Prompt tracking for questions GSC does not store is Promptwatch.

The three official URLs

AI features and your website. Eligibility. Overviews and AI Mode. Fan-out, with Google's own lawn-weeds example. Supporting links need to be indexed and snippet-eligible. No extra technical requirements beyond ordinary search eligibility. Preview controls are nosnippet, data-nosnippet, max-snippet, and noindex. Google-Extended is not a clean Overview-only off switch, so a team that toggles it expecting to keep Search and lose Overviews will be surprised on both ends.

The eligibility page is the one that governs what gets in, and the eligibility page is the one that says no extra technical requirements beyond ordinary search eligibility. That sentence is the one that kills the most invented factors, because a vendor who bills a special Overview technique has to square the technique with a page that says no extra requirements exist. The Google-Extended point is the one that prevents the most accidental opt-out, because a team that toggles Google-Extended to leave Overviews and keep Search toggles a control that does not do that, and the surprise is on both ends, meaning the team loses Search too.

Optimizing your website for generative AI features on Google Search. Writing and site work. Unique point of view. Non-commodity pages. Headings for humans. Structured data that matches the visible page. Schema is not required for Overviews, so a bill for "special Overview schema" is a bill for something Google did not ask for. Ignore llms.txt for Google Search. Do not publish a thin URL for every fan-out phrasing, because doorway pages are still doorway pages. AI-assisted drafts still pass Search Essentials and spam policy. Helpful content was not replaced by the AI guidance.

The optimization guide is the one that governs what to write, and the guide is the one that dismisses the two most common invented tactics. The special Overview schema bill is a bill for something the guide says is not required, which is the part that makes the bill a bill for nothing. The llms.txt tactic is a tactic the guide says to ignore for Google Search, which is the part that ends the argument when a stakeholder asks for it. The doorway-page warning is the one that prevents the most common fan-out misread, because a fan-out is a cluster of questions and a thin URL per phrasing is a set of doorway pages, and doorway pages are still doorway pages regardless of the AI context.

Generative AI performance reports (June 2026). Impressions in Overviews, AI Mode, and Discover generative features. Pages, countries, devices, dates. Clicks are not in that list, which is the gap people trip over. Mixed into the main Performance report, Web, as well.

The performance reports announcement is the one that governs what you can measure, and the announcement is the one that lists the dimensions and the metrics. The dimensions are pages, countries, devices, and dates. The metric is impressions. Clicks are not in the list, which is the gap that produces the most common reporting error, because a stakeholder who reads "impressions" as "clicks" reads a visibility number as a traffic number, and the two are different. The impression is the appearance in the Overview, and the click is the session that followed, and GSC stores the first and not the second for the generative surfaces.

If a vendor PDF disagrees with those three, we keep the three. The vendor PDF is not the source, and a slide built on it will not survive a question about where the claim came from.

The vendor-PDF point is the one that keeps the documentation as the authority, because a vendor PDF is a marketing artifact and the documentation is the spec. A slide built on the PDF is a slide that falls apart the first time a stakeholder asks for the source, because the source is a vendor and not Google. A slide built on the documentation is a slide that survives the question, because the source is Google and Google is the authority on its own surfaces.

What the documentation does not cover

ChatGPT, Claude, Perplexity. Bing Copilot, which is covered by Bing AI Performance. The prompt that triggered an Overview. The other domains in the module. These are the gaps that make a second tool necessary, because Google's own reports stop at the impression and do not name the competing links.

The gaps are the ones that decide whether a second tool is needed, because the documentation covers Google surfaces and the gaps are the non-Google surfaces. ChatGPT, Claude, and Perplexity are not Google surfaces, and Google does not document them. The prompt that triggered an Overview is a unit Google does not store, which is the gap that makes prompt tracking a separate tool. The other domains in the module are the competing links, which Google does not name, which is the gap that makes citation analytics a separate tool. Each gap is a piece of the picture the documentation does not hold, and the gaps together are the case for a second tool.

We measure that gap in Promptwatch on paid plans, covering ChatGPT, Gemini, Claude, Perplexity, and Overviews daily. Essential is $95 a month. Kick-off is $199 a month. Professional is $245 a month. Business is $579 a month. Explore is ChatGPT only with 10 prompts, so it is a first look rather than a monitoring tier. Ahrefs Brand Radar and Semrush AI Toolkit, at $99 a month per domain, are modeled rather than fetched. They are not Search Central. Keep Ahrefs and Semrush for ordinary SEO, where they belong, and keep the documentation as the Google authority.

The modeled-versus-fetched point is the one that separates the SEO suites from the tracker, because modeled data is a projection and fetched data is a record. A modeled mention is the model's idea of what the engine would write. A fetched mention is the engine's actual output. The record is the stronger signal, and the SEO suites that model the data are not the same kind of tool as the tracker that fetches it. The SEO suites belong in the ordinary SEO stack, where their modeling is a useful extra eye, and the documentation belongs as the Google authority, which is the thing the SEO suites are not.

How we use the documentation on an account

Put both Search Central URLs in the project brief. Anyone proposing llms.txt as a Google tactic has to quote the optimization guide back, which usually ends the argument.
Audit money URLs against the eligibility page: index, snippet, robots, CDN. Technical tickets come from our audit checklist.
Audit a sample of pages against the optimization guide: first-party point of view, no doorway fan-out URLs, markup that matches the offer.
Turn on GSC generative reports. Train stakeholders on "impression," because the word is not a click and the confusion is the most common reporting error.
Only then decide whether prompt tracking is in the SOW. Documentation-only retainers can stop at GSC, and that is a legitimate scope rather than a lost deal.

The five steps are the order that keeps the documentation in charge of the program, because each step uses a specific official page for a specific job. The brief uses the URLs, which is the step that makes the documentation the reference. The eligibility audit uses the eligibility page, which is the step that catches the technical gaps. The content audit uses the optimization guide, which is the step that catches the writing gaps. The GSC step uses the performance reports, which is the step that sets the measurement. The SOW step is the one that decides whether a second tool is needed, and a documentation-only retainer is a legitimate scope because not every program needs a tracker.

We will not bill special Overview schema. We will not bill synonym-chasing rewrites the guide dismissed. We will not claim Google-Extended turns Overviews off and leaves Search on, because the documentation says it does not.

The three will-nots are the three invented tactics that the documentation dismisses, and the will-nots are the ones that keep the retainer honest. Special Overview schema is a bill for something the guide says is not required. Synonym-chasing rewrites are a bill for something the guide dismisses. The Google-Extended claim is a claim the documentation contradicts. Each will-not is a tactic that produces a bill without a basis in the documentation, and the will-not is the discipline that keeps the bill tied to the source.

FAQ

Is there an official AI Overviews ranking-factor list?

No. Eligibility plus ordinary SEO plus people-first pages.

Which document wins if they seem to overlap?

Eligibility and controls: the AI features page. What to write and what to skip: the optimization guide. What you can measure in GSC: the June 2026 announcement.

Do we implement IndexNow because Search Central mentioned AI?

IndexNow is a Bing and participating-engine protocol (indexnow.org). Search Central's AI pages are Google. We may still enable IndexNow. We will not cite it as Google documentation.

What to do this week

Open the three official URLs. Bookmark them.
Check money URLs for indexation and snippet controls.
Open GSC generative reports.
If the SOW includes ChatGPT or Perplexity, add Promptwatch. If it is Google only, stay on the documentation.
Email hello@1001seomedia.com if you want the documentation mapped onto your URLs. We follow Search Central. We do not replace it with a tool blog.