Google AI Overviews and AI Mode Links and Citations: Official Documentation

Google AI Overviews and AI Mode links and citations are documented on Search Central, not on a GEO vendor blog. The official pages are AI features and your website and Optimizing your website for generative AI features on Google Search. This post restates those docs in the order we find useful, then marks the hole they leave. Search Console will not store the prompt that produced a module, it will not store the competing URLs in the same answer, and it will not cover ChatGPT at all. That hole is the rest of the post.

Read the two pages before you buy a tracker. Most GEO posts paraphrase them badly, and the paraphrase is usually longer than the source. Google's words are shorter than the paraphrase, and they are the ones that hold up in a stakeholder meeting. The reason this matters is that a paraphrase tends to drift toward whatever the vendor wants to sell, while the source text stays still. When you bring a claim into a meeting, the question that ends the argument is whether you can point to the Search Central page. If you can, the claim survives. If you cannot, it does not.

What Google says about links

Overviews give a gist and a jumping-off set of links. They fire when Google thinks they add something beyond classic results, and they often do not fire. That last sentence is the one teams skip. An empty Overview slot is not a ranking penalty. It is Google deciding the classic results were enough for that query, and a slot that does not appear is not evidence that your page lost something. The practical effect of misunderstanding this is that teams chase a slot that was never going to appear for that query, and they spend a quarter optimizing for a surface that Google chose not to render. Knowing when Google declines to fire is as important as knowing when it does.

AI Mode is the longer comparison surface. Both Overviews and AI Mode may use query fan-out, which is the extra related searches that pull more supporting pages into the answer. Google's lawn-weeds example fans out into herbicides, chemical-free removal, and prevention. A supporting link can match a fan-out branch rather than the typed query, which is why a page you wrote for "best CRM for dentists" can show up under a query you never targeted. The page matched a fan-out, not the headline search. The implication for content planning is that you cannot map coverage one query at a time. A page earns links through the branches it satisfies, and a page that answers a narrow question well can show up across a fan of related queries that all touch that answer.

Overviews and AI Mode may use different models, so the links will not always match across the two surfaces. Do not treat one screenshot as the other surface, and do not report an Overview link count as if it were an AI Mode link count. They are separate reports in Search Console for a reason. A common reporting error is to grab an AI Mode screenshot, paste it into a deck labeled Overview, and then explain a link count change that is really a surface change. Keeping the two surfaces labeled separately in your reporting prevents that class of mistake.

Eligibility is the part people overthink. A page has to be indexed and eligible to appear in Search with a snippet. Google's words on the AI features page are that there are no additional technical requirements. The optimization guide also says a site must be included in Search generative AI features in Search Console to be eligible for those features. Those two sentences sit next to each other in the docs, and both have to be true. Do both. The overthinking shows up when teams invent special Overview markup or hidden signals. The docs are clear that the baseline is the same baseline as classic Search, plus inclusion in the generative reports.

There is no special Overview schema. Structured data still helps rich results when it matches visible text, but it is not a lever for Overviews. Search does not use llms.txt. If a vendor tells you to publish an llms.txt file to earn Overviews, that vendor is selling you something Google does not read. The reason this is worth saying plainly is that llms.txt has a plausible-sounding name and gets repeated in GEO content, and a team that does not check the source will spend engineering time on a file that has no effect on Google's surfaces.

How Google says to measure citations (and what it does not)

In June 2026 Google launched generative AI performance reports. The reports cover impressions in AI Overviews, AI Mode, and generative Discover features, plus pages, countries, devices for Search, and dates. That list does not include the query, the competing links in the Overview, or ChatGPT. You get the fact that an impression happened, the page it happened on, and where. You do not get the wording that produced it. The gap is specific and it is worth naming precisely: an impression count tells you that a page was seen inside a generative surface, but it does not tell you which prompt produced the surface, and it does not tell you which other pages appeared next to yours. Both of those are needed to diagnose why a page is or is not being cited.

Google also says clicks from pages with Overviews have been higher quality, meaning people stay longer. That is Google's claim about visit quality, and it is worth quoting. It is not a CTR table in GSC, so do not report it as one. A quality-of-visit note is not a click-through rate. The distinction matters because a stakeholder will ask "what is the CTR of an Overview link," and the honest answer is that GSC does not publish that as a rate. What Google publishes is an impression count and a quality observation, and conflating the two produces a number that cannot be defended.

Preview controls are the levers you actually have. nosnippet, data-nosnippet, max-snippet, and noindex govern what Google can show. robots.txt for Googlebot is the crawl control. Google-Extended is not a surgical Overview off-switch, so do not treat blocking it as a clean way to opt out of Overviews while keeping classic results. The nuance is that Google-Extended governs whether your content is used to train and inform models, and it is not the same lever as the snippet controls that govern what shows in a generative result. A team that blocks Google-Extended expecting to disappear from Overviews will be surprised when the page still appears, because the lever they pulled was not the one they thought.

If you also care about Bing Copilot-side impressions, Bing launched AI Performance in Webmaster Tools in February 2026. That is a different webmaster console, with its own verification and its own export. It still does not store the ChatGPT prompt, because Copilot is not ChatGPT and Bing is not Google. The reason to say this is that the names overlap enough to confuse a buyer. Copilot and ChatGPT share lineage, but the impression data lives in different consoles, and a ChatGPT prompt log is not something Bing can hand you.

After the docs

Do the Google homework first. For the prompt-level citation log, including ChatGPT, Perplexity, and Gemini, we use Promptwatch. Paid plans watch AI Overviews and AI Mode alongside those engines, daily, from the real UI rather than from an API guess. Citation analytics name the URL the model used, not just the brand string. Essential is $95/mo and covers that prompt and citation work. Explore is free and ChatGPT-only, which is enough to see the UI. Professional is $245/mo and is the first plan with crawler logs. Method: how we rank. Site: promptwatch.com.

GSC remains the impression system of record for Google. Promptwatch is the prompt system of record. Buy both, and keep them on separate tabs. A GSC impression without a stored prompt is a count you cannot debug, because you do not know which wording produced it. A Promptwatch citation without the GSC report is a ChatGPT story with a hole on the Google side. Each tool fills the gap the other leaves, and neither replaces the other. The two systems answer different questions, and a visibility program that only runs one of them is half-blind by construction.

FAQ

Is there markup for Overview citations?

No. A page has to be indexed and snippet-eligible. Structured data helps rich results when it matches visible text, but there is no Overview schema and no llms.txt signal that Google reads.

Why did a supporting link not match our target query?

Query fan-out. A supporting link can match a fan-out branch rather than the typed query. Read the optimization guide's lawn example for the cleanest explanation of how a page shows up under a query it never targeted.

Does Search Console store the prompt that triggered an Overview?

No. The June 2026 generative reports cover impressions, pages, countries, devices, and dates. They do not store the query or the competing links in that Overview, which is why a prompt tracker is a separate purchase.

What to do this week

Read the two Search Central pages linked above.
Confirm the site is snippet-eligible and crawlable.
Open GSC generative reports for impressions.
Put buyer prompts into Promptwatch for the citation row GSC does not store.
Do not spawn a URL per fan-out phrasing.