LLM Visibility Tracking: How to Measure a Brand Across Models

The ChatGPT your customer uses is not one model. Secondary reports as of September 2026 said Free and Go users got gpt-5.6-luna, Plus users got gpt-5.6-sol, and higher tiers got a GPT-6 Pro model. OpenAI has not published that mapping itself.

OpenAI does publish knowledge cutoffs. The GPT-5.6 family has a Feb 16, 2026 cutoff. The Instant path, which OpenAI calls chat-latest, has an Aug 31, 2025 cutoff. So "what ChatGPT knows about you" has at least two answers, depending on routing.

A tracker calls an API. It pins one model, one set of settings and one account state. That makes results repeatable, which is the point. It also means the tracker measures a stand-in for what buyers see. Logged-in users add another layer: OpenAI says ChatGPT may use saved memories when it rewrites a search query. Two buyers asking the same question can get different answers.

What this means for you: read tracking numbers as a controlled sample of model behavior. The gap between that sample and real sessions is covered in how API answers differ from consumer app answers. The next move is to decide which consumer experience each tracked model is standing in for, and write that down.

An answer with live web search and an answer without it come from different stages of the system.

• Retrieval off (parametric). The model answers from what it learned before its cutoff. Nothing you publish this quarter reaches this answer.
• Retrieval on. The model fetches pages for the query. Crawler access, indexing and your current pages now matter.

The apps decide per query. As of October 2026, ChatGPT search runs at the model's discretion. Claude.ai web search is a user toggle with no published default. APIs differ again. Perplexity's Agent API requires you to opt in to web search, and code left on the retired Sonar endpoint after 2026-09-27 can return ungrounded answers without any error. Check whether AI model APIs search the web by default before trusting any tool's numbers.

Nobody agrees how often ChatGPT searches. Profound reports that about 18% of conversations trigger at least one web search. Semrush's US clickstream data put it at 34.5% of queries in February 2026. cloro reported 80.5% globally on a 100-prompt set it chose. These measure different things on different populations, so none of them tells you what happened to your brand's answer.

The rule: if you did not see citations, and you did not control the setting, you do not know whether the answer was grounded. Record retrieval state on every answer, or do not compare answers at all.

Product names hide the split between surfaces. Google is the clearest case. It runs at least four generative paths:

• The Gemini app.
• The Gemini API with Grounding with Google Search, which developers call and pay for separately.
• AI Overviews in Search.
• AI Mode in Search.

Google states that AI Overviews and AI Mode "may use different models and techniques" and can return different links. AI Overviews also often do not trigger at all. A tool that calls the Gemini API with grounding is measuring a proxy for the Gemini app. It is not measuring AI Overviews or AI Mode. AI Overviews has no official API, so the practical route is SERP data that captures the overview Google actually showed for a query. See whether you can track what Google AI Mode says for that surface.

Perplexity is similar. Secondary reports as of September 2026 said free users get an auto-router called "Best" and paid users pick from a menu of models. No single API model is the app.

What this means: label each tracked column with the surface it stands for, not the brand. "Gemini API, grounded" is honest. "Google AI" is not.

In your control

• Which buyer questions you track, and in which countries and languages.
• Which model IDs you call, and whether retrieval is on.
• Account state: logged-out, memory off, no custom instructions.
• When you run, and how often.
• On your own site: per-bot robots.txt permission, firewall and CDN reachability, server-side rendering of key facts, and stating specific facts (prices, numbers, one-line definitions) on pages crawlers can fetch.

Out of your control

• Which model a buyer's plan routes them to.
• Model releases and silent updates. OpenAI's model docs warn the snapshot "will be regularly updated."
• How the engine rewrites the question. OpenAI says ChatGPT search typically rewrites a query into one or more targeted queries.
• Whether retrieval fires, which pages enter the context, and which get cited.

The evidence on content tactics is thin. A critical survey of 45 GEO studies (arXiv, July 2026) found no technique with a stable, longitudinal, cross-platform causal effect on discoverability. So use tracking to find problems you can fix, like a blocked crawler or a wrong price, not to grade tactics. Expect answers to shift when vendors ship, as covered in whether a new model release changes brand answers.

Every answer you store should carry enough context that someone else could rerun it. Here is the minimum record.

In a report, put a one-line disclosure under every chart. A hypothetical example: "ChatGPT column: gpt-5.6-luna via API, web search on, logged-out, US English, runs of 1 to 3 October 2026. Approximates ChatGPT Free."

If a number blends retrieval-on and retrieval-off answers, say so. If it blends platforms, say that too, and show each platform underneath.

1. Write the question set. Use real buyer questions, and decide whether to track branded prompts, unbranded prompts, or both. Keep the wording fixed between runs.
2. Pick surfaces and pin models. For each surface, record the model ID and which consumer plan it approximates.
3. Run each question with retrieval off and on. The off answer shows what the model believes from training. The on answer shows what it says when it can read the web.
4. Record retrieved and cited sources separately. A page that is retrieved but not cited tells you access works. A page that is never retrieved points you to access and indexing checks.
5. Check facts, not just mentions. A mention with the wrong price is a problem, not a win. Compare each claim against your own verified facts.
6. Add engine-reported data. Bing Webmaster Tools' AI Performance report shows grounding queries, the queries the AI actually ran. Google Search Console's generative AI report shows impressions only. In Otterly's 2026 experiment, Microsoft's report credited one site with 42,817 citations while Otterly's prompt tracker logged 3,934, though the time windows may not match. Read more on prompt-based tracking vs engine-reported data.
7. Re-run on a schedule, and after model changes. Set a fixed cadence, as in how often to re-run AI visibility checks, and log every model change next to the trend line.

If you want a starting point before building any of this, the free checker on this site runs up to three buyer questions as a one-off snapshot. Record its date like any other reading.

• One blended score across platforms. In Otterly's experiment, Microsoft Copilot produced 2,718 of one site's 3,934 tracked citations while ChatGPT, Gemini, AI Overviews and AI Mode almost never cited it. A blended score would have hidden that.
• Reporting an AI rank or position. Google's generative AI report has no position data, so no position number can be checked against ground truth. Report presence, accuracy and citations.
• Comparing across a model change without a note. A drop after a vendor release may have nothing to do with your site.
• Treating cited as used. Citation counts are a noisy proxy for influence. Research on how models cite found citation behavior can be changed without changing the sources.
• Testing while logged in. Memory and custom instructions make results personal and unrepeatable.
• Calling the Gemini API "AI Overviews". They are different products.
• Blaming your site for a retrieval-off miss. If search was off, the answer came from training data, and site fixes cannot reach it this quarter.