Brand mentions and citations in Gemini answers

The short answer: With the Google Search grounding tool switched on, Gemini decides whether to search, rewrites your question into its own search queries, and writes a reply from what came back, attaching each citation to a character range of its own sentence. That is the flow Google documents for the Gemini API. So in a grounded answer your brand turns up when a page the model fetched says the thing, in words that survive being lifted out of context, and retrieval is doing more of the work than training data is.

How brand mentions in Gemini work, and how they differ from brand citations. They are two separate events inside the same reply. The mention is your name appearing in the answer text. The citation is a url_citation annotation binding a source to a character range of that sentence. A grounded reply can name you without citing you, and can cite your page in a sentence that names somebody else as the recommendation, so decide which of the two you are counting before you report a number upward.

That sounds like a small distinction. It changes what you fix first.

The loop, in the field names Google publishes

You do not have to guess at this part. Call the Gemini API with the google_search tool enabled and the response comes back as an ordered list of steps, documented field by field. Whether the Gemini app runs the same path is not visible from outside Google, so read what follows as the API's behaviour. There is a thought step. Then a google_search_call step holding a queries array, which is the model's own rewrite of what the user typed. Then a google_search_result step. Then the answer itself, with url_citation annotations that carry a url, a title, a start_index and an end_index.

Read that last part again, because it is the whole game. The citation is not attached to the answer. It is attached to characters 0 through 56 of the answer. Gemini is asserting that a specific source supports a specific clause, and the clause is usually one sentence long.

Two things follow. First, the queries the model invents are not the question the user asked, so optimising for the user's phrasing is a guess at an intermediate step you cannot see. We pull the queries array during audits because it is the clearest view we can get of what was actually searched. Second, if a citation has to bind to one clause, a page that supports half a sentence is a better candidate than one that gestures at the whole topic. That is an inference from the documented response format, not something we have measured. How each AI engine picks the sources it cites covers how differently the other engines handle that same step.

Gemini is not AI Overviews, and one audit measured the gap

People treat the two as one surface because Google owns both. One published audit argues against that. A preprint posted on 16 September 2026, auditing AI-generated product recommendations, curated 2,528 real commercial-advice queries and evaluated 1,536 responses from ChatGPT, Gemini and AI Overviews. For the same query it reports the ChatGPT and Gemini interfaces sharing 5.4 percent of displayed domains on average, with no domain in common in 76.7 percent of comparisons. It separately reports that each assistant's API exposes a different source set from its own interface, with mean domain overlaps of 12.0 percent for ChatGPT and 14.8 percent for Gemini.

The same preprint reports a first-person product preference in 7 percent of Gemini's product-recommending responses, against 79 percent for ChatGPT and 2 percent for AI Overviews. In that dataset Gemini reported more than it endorsed. This is one audit of one vertical, shopping advice, and it has not been peer reviewed, so we would not carry the figures across to your category untested. It is still worth asking whether your pitch leans on enthusiasm a given surface may not supply.

None of this means AI Overviews work is wasted. It means the two need separate measurement, and how Google AI Overviews chooses its sources is a different investigation from this one.

The control that quietly removes you

Here is a failure mode we keep finding unchecked, and it is a one-line fix.

Google documents a robots.txt product token called Google-Extended. Its crawler documentation, last updated 14 July 2026, says the token has no separate HTTP request user agent string and that crawling is done with existing Google user agent strings, so you should not expect to see Google-Extended itself in your access logs. The same page says the token controls whether content Google crawls may be used for training Gemini models and for grounding, which it defines as providing content from the Google Search index to the model at prompt time, in Gemini Apps and in Grounding with Google Search on Vertex AI. It also states that Google-Extended does not impact a site's inclusion in Google Search and is not used as a ranking signal there.

Sit with the asymmetry. On Google's own account the block does not cost you inclusion or ranking in Search, and Google's AI features guidance points site owners to robots.txt for Googlebot and to snippet controls for what Search can show, treating Google-Extended as the lever for its other systems. What the block can do is remove your pages from grounding in Gemini Apps. We have found it still in place on client sites during audits, usually added years earlier through a plugin or a CDN rule. We have not sampled the web, so we cannot tell you how common it is. A team watching healthy AI Overviews impressions has little reason to go looking.

The check takes about forty seconds:

`` curl -s example.com/robots.txt | grep -i -A3 "google-extended" ``

If a Disallow: / sits under that token, somebody chose it. Find out who and when, then decide again with the 2026 facts rather than the 2023 ones. Separately, Google's AI features page says nosnippet, data-nosnippet, max-snippet and noindex limit the information shown from your pages in Search. We read that as also capping the passage an AI feature in Search can lift, though Google does not say so in those words. Three different levers, three different owners, and often nobody holding the map.

What Google added for site owners in 2026

On 3 June 2026 Google announced a Search Console toggle letting an owner decide whether a site appears in and helps ground responses in generative AI Search features, alongside impression data showing which pages turn up in AI responses and in which countries. It began with a subset of website owners in the UK, and a footnote on the same post, updated 31 August 2026, says the features reached all websites worldwide on that date. The announcement put AI Overviews at more than 2.5 billion monthly active users and AI Mode above one billion.

Two cautions. That toggle governs AI features inside Search, which is not the same surface as the Gemini app, so it does not replace the Google-Extended check above. And impressions in an AI response are a presence metric, not a mention metric. Your page can be a cited source for a paragraph that names a rival as the recommendation. Worth knowing which you are counting before you report it upward.

Writing for a citation that lands on one clause

Since the annotation binds to a character range, the useful question is narrow: which single sentence on your page would a model be willing to stand behind.

Put the claim and its qualifier in the same sentence. "Our starter plan is 500 dollars and covers one audit cycle" can be lifted whole as one span. A price in a table three screens below a paragraph that says pricing is flexible cannot. We are reasoning from the documented span structure here, not reporting a measured result.

Keep the same fact worded the same way across your site. We cannot see how a model resolves two versions of your pricing or your founding year, and neither can anyone outside the vendor. What we can say is that a hedged mention reads to a buyer much like no mention, so removing the contradiction is cheaper than finding out.

Write to the questions, not the phrases. The model is inventing its own queries anyway, which is the practical case made in prompts are not keywords: planning content for questions.

Then test it the slow way. Ask the same brand question several times across a few days, log whether you were named, whether the claim beside your name was true, and which domains got cited. One run is an anecdote. The shopping-advice preprint above found the products recommended often changed across repeated requests, which is the argument for logging a series rather than a single reply. The self-check is ten of these you can run this afternoon without buying anything.

Honest limits

We do not know how closely the public Gemini app matches the API grounding path: the field names above are documented for the API, and the app carries retrieval, ranking and safety layers nobody outside Google can inspect, so treat the API as a good model of the app rather than a copy of it. The overlap and preference figures here come from a single preprint on shopping advice, posted 16 September 2026, not peer reviewed and not run on your category. We have found stale Google-Extended blocks on client sites during audits, but we have not sampled the web and cannot say how common they are, and we do not publish a count we cannot reconstruct. The writing guidance in this post is reasoned from the documented citation span rather than from a controlled test, which we have not run. We also cannot give you a date: access fixes such as a robots.txt line can clear within a crawl cycle, while anything depending on third parties repeating your claim moves on its own schedule, which is the argument in how long AEO takes to show up in answers.

FAQ

Does Gemini answer brand questions from training data? With the search tool on, Google's API documentation shows the response carrying the search queries the model ran and the URLs it cited, which means that answer was assembled from fetched pages. Ungrounded replies do exist and they are the ones most likely to be stale or wrong about you. We cannot tell you what share of real Gemini app sessions are grounded.

Is blocking Google-Extended the same as blocking Googlebot? No. Google's documentation states that Google-Extended does not affect a site's inclusion in Google Search and is not a ranking signal. It governs training and grounding in Gemini Apps and in Grounding with Google Search on Vertex AI.

Why does Gemini cite different sources than AI Overviews? Different retrieval and different ranking, inside one company. The September 2026 shopping-advice preprint cited below measured low domain overlap between assistant interfaces on identical queries. That is one audit of one vertical, so treat a strong result on one surface as weak evidence about another rather than as settled.

How do brand mentions work in Gemini? With grounding on, your brand is named when a page the model fetched says the thing in words that survive being lifted into one sentence. The mention sits in the answer text. The citation is a separate url_citation annotation tied to a character range of that sentence, so a reply can mention you without citing you and cite you without mentioning you. Google does not publish how often either happens in the Gemini app.

How do I track brand citations in Gemini? Through the API you can read both directly: the response carries the model's own queries array and url_citation annotations with the URL and title behind each span. For the Gemini app Google publishes no equivalent report, and the Search Console AI impression data covers AI features inside Search rather than the app, so the honest method there is to run the same question repeatedly across a few days and log what was named and what was cited.

How many prompts do I need before the numbers mean anything? More than you want to run by hand. A single result is an anecdote because the model rewrites the query each time. Our audits use a fixed set of about 45 prompts across five engines, run with retrieval on and off, so a finding can be argued with rather than believed.

Sources

All five were opened and checked on 12 October 2026.

• Google, Gemini API documentation, grounding with Google Search: response steps, google_search_call with its queries array, and url_citation annotations carrying start_index and end_index. ai.google.dev/gemini-api/docs/google-search
• Google, list of common crawlers, Google-Extended section, page last updated 14 July 2026: no separate request user agent string, scope over training and grounding in Gemini Apps and Vertex AI, no impact on Search inclusion or ranking. developers.google.com/crawling/docs/crawlers-fetchers/google-common-crawlers
• Google Search Central, AI features and your website, page last updated 10 December 2025: nosnippet, data-nosnippet, max-snippet and noindex limit the information shown from your pages in Search. developers.google.com/search/docs/appearance/ai-features
• Google, new opportunities, control and insights for website owners, posted 3 June 2026, updated 31 August 2026: the Search Console toggle, the AI impression insights, the UK test, the worldwide rollout footnote, and the 2.5 billion and one billion monthly user figures. blog.google/products-and-platforms/products/search/new-controls-website-owners/
• Auditing AI-Generated Product Recommendations, arXiv preprint 2609.18729, submitted 16 September 2026, not peer reviewed: 2,528 commercial-advice queries, 1,536 evaluated responses, the 5.4, 76.7, 12.0 and 14.8 percent domain figures and the 79, 7 and 2 percent first-person preference figures. arxiv.org/abs/2609.18729