Tracking What Gemini Says About Your Brand

Google runs more than one generative product, and each needs its own tracking. The Gemini app is what consumers type into. The Gemini API with Grounding with Google Search is a separate, developer-invoked path. AI Overviews and AI Mode sit inside Google Search. Google's own documentation says AI Overviews and AI Mode "may use different models and techniques" and can return different links.

The next move: decide which surface you are reporting on before you write a single prompt. If AI Mode matters to you, read whether you can track Google AI Mode separately. A score labelled "Gemini" that blends these surfaces tells you nothing you can act on.

Most tracking tools, including any script you write, call the Gemini API. The app is a different product. In a snapshot of vendor pages taken on 21 September 2026, Gemini's release notes listed gemini-3.6-flash in the app's model menu, while the API had already shipped gemini-3.8-flash. Google did not state which model free app users get by default.

That gap matters in three ways:

• Model version. A newer or older model can know different facts about your brand from its training data.
• Search behavior. The app decides for itself when to search. Your API call only searches if you enable grounding.
• Defaults that move. Model menus and defaults change with release notes, and nothing alerts you.

The broader pattern is covered in API answers vs consumer app answers. The next move: label API results as "API proxy for the Gemini app", pin the exact model ID in every run, and re-check your mapping whenever Google's release notes or API changelog change, as of October 2026.

Retrieval is per query and discretionary. With Grounding with Google Search enabled, Google's Gemini API documentation says the model "automatically generates one or multiple search queries" and runs them. You do not choose those queries, and you cannot see the ones the app runs.

This splits every answer into one of two kinds:

• Ungrounded (parametric). The model answers from training data frozen at a knowledge cutoff. Nothing you publish this quarter changes it.
• Grounded (retrieval). The answer depends on what Google Search returned for queries the model wrote.

Averaging the two hides which problem you have. A wrong price in an ungrounded answer is a training-data issue. A wrong price in a grounded answer often points at a page the model read.

Log these fields for every prompt:

1. Date, model ID and grounding setting.
2. Whether the response actually came back with grounding data.
3. The citations returned, and the search queries where the API returns them.
4. Whether your brand was named, and which facts were stated.
5. Whether each fact was correct against your own records.

Claude has the same split through a user search toggle, covered in tracking Claude's answers.

Google's Gemini API documentation says grounding is billed per search query on Gemini 3 models and per prompt on Gemini 2.5 and older, as of October 2026. Because one prompt on a newer model can generate several searches, the cost of a prompt is not fixed. Two identical runs can cost different amounts.

This pushes you toward fewer prompts than you would like. Small samples swing. Say you track a handful of buyer questions once a month: one odd answer can move your result a long way, and you cannot tell noise from change.

What to do about it:

• Fix a prompt set and keep it stable, so runs are comparable.
• Log the billed cost of each run, then size the set to your budget.
• Report rates over answers that returned, never over answers requested. A failed call is missing evidence, not a clean answer.
• Repeat the same prompt more than once before you treat a change as real.
• Never compare a run on one model version with a run on another as if it were a trend.

Some levers are documented. Most of what GEO vendors sell is not.

Documented controls

• Google-Extended. Google's crawler documentation says this token controls use of your content for Gemini Apps training and grounding, and that it "does not impact a site's inclusion in Google Search nor is it used as a ranking signal." Check that your robots.txt does not contain a block you did not intend:

User-agent: Google-Extended
Disallow: /

• Index and snippet eligibility. Google says a page must be "indexed and eligible to be shown in Google Search with a snippet" to appear as a supporting link in AI Overviews or AI Mode. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out. No robots.txt token blocks AI Overviews or AI Mode specifically.
• Rendering. Googlebot renders JavaScript, so client-side rendering is less of a risk on Google surfaces. It still hides content from OAI-SearchBot, ClaudeBot and PerplexityBot.

What has some support

Stating a specific fact (a price, a number, a one-line definition) on a fetchable page gives an engine a sentence it can quote. Being named on the sites Gemini cites instead of yours is associated with appearing in answers, with no promised effect.

What is not a lever

Google's AI optimization guide says you don't need new machine-readable files, AI text files or special markup for its generative features, and you don't need to write in a special way for AI.

Google documents a fair amount about access and billing. It documents very little about selection.

Documented: how grounding generates queries, how grounding is billed, what Google-Extended governs, and the eligibility rules for AI Overviews and AI Mode.

Not documented: which model free Gemini app users get by default, how grounded answers choose sources, or any position inside an answer. Search Console's generative AI report covers AI Overviews and AI Mode with impressions only, so check whether Search Console reports Gemini app citations before relying on it.

The research does not close this gap either. A July 2026 critical survey of 45 GEO studies (arXiv 2607.14035) found no technique with a stable, longitudinal, cross-platform causal effect on discoverability.

The next move: report presence and factual correctness per surface, with model and grounding state shown next to every number. For a first look, the free checker on this site runs up to three buyer questions as a snapshot, and you can start with a free check of what Gemini says about your brand. The same logging discipline applies to monitoring ChatGPT.