How to Track Your Brand’s Visibility in ChatGPT and AI Search

You can track brand visibility in ChatGPT, and in the other assistants your buyers use, by running the same buyer prompts regularly, saving the full answers, and measuring whether your brand is mentioned, recommended, cited, or omitted.

Define visibility before you measure it

Brand visibility in ChatGPT isn't one score. Measure four separate outcomes: whether your brand is mentioned, whether it is recommended, whether the answer gives a reason to choose it, and whether the response links to a page you control.

That distinction matters because a brand can appear in an answer without being a serious option. A model might list your company among several providers, describe it incorrectly, or mention an outdated page. Record each outcome separately instead of marking the test as simply visible or invisible. If you want the wider picture first, of why these answers now matter as much as a search ranking, start with what AI visibility is.

Use a small tracking sheet with columns for prompt, platform, model if shown, date, location, brand mention, recommendation, stated reason, linked source, competitor mentions, and factual errors. Add a final column for buyer usefulness. A response that names your company but gives the wrong service area is not a successful result.

Treat the sheet as a measurement record, not a prediction of every customer conversation. ChatGPT responses can vary with wording, conversation history, account settings, location, and whether web search is available. Your result shows what happened under a defined test condition.

Choose the buyer prompts to test

Start with the questions a ready-to-buy customer would actually type, not prompts that contain your brand name. Pull them from three places: the questions prospects ask on sales calls, the questions that land in your support inbox, and the plain-language requests a buyer would type into an assistant before they know your name.

Cover four kinds of question. A category question asks who serves a need. A shortlist question asks for several suitable providers. A comparison question puts your kind of business beside an alternative, for example "alternatives to [a competitor you know]". A problem question asks what kind of specialist could solve a specific situation. For example, a freelance bookkeeper could test, “Who can help a small business in Manchester catch up on overdue accounts?”, “What should I look for in a bookkeeper for a growing consultancy?”, and “Compare local bookkeeping services for a business that needs monthly reporting.” Replace the details with your market, service, location, budget, and urgent problem.

Use the words a buyer knows, not the words your website uses. If customers say “bookkeeping help”, don't test your internal phrase “financial operations support”. And write every prompt down before testing any assistant: if you quietly improve the wording after a weak answer, measurement has turned into persuasion.

Keep one core set of six to ten prompts unchanged from check to check, and a separate set for experiments with a new service or market. The important test is not whether ChatGPT can mention you after being led to your name. It is whether a buyer can reach you from a plausible question.

Run a repeatable check without code

Run a manual prompt panel in a fresh chat, using the same wording, account type, location, and browsing setting wherever possible. Open ChatGPT, paste one saved buyer prompt, and save the complete response with the date and visible model or search indicators. Then run the same prompt in the other assistants you track, each in its own fresh session.

Use a spreadsheet or a shared document for the record. Give each prompt its own row and each platform its own column, or store one row per prompt-platform-date combination. Copy the answer rather than recording only a yes or no. The surrounding wording reveals whether the mention was prominent, qualified, or wrong.

Repeat each important prompt in a new conversation. A previous conversation can teach the model your preferred answer and make a later result look stronger than an ordinary buyer would receive. Avoid switching between personalisation settings, logged-in accounts, and locations without recording the change. If an assistant asks for a missing detail such as your city, save its first answer, then run a second, clearly labelled version with the detail supplied rather than silently editing the original prompt.

A repeatable check does not require an API, automation platform, or technical setup. If you later automate collection with a tool such as Zapier, Make, or n8n, keep a manual sample as a control. Automated runs can preserve consistency, but they cannot make a changing model response objectively stable.

Record the evidence from each answer

Record the answer’s evidence, not just the brand name. Capture the exact prompt, full response, date, model label, whether web search was used, cited URLs, and the position or prominence of your mention.

Then classify the result using consistent labels. “Absent” means no mention. “Listed” means your brand appears without a clear preference. “Recommended” means the answer gives a reason to choose you. “Cited” means the response links to a page that supports the claim. “Found but wrong” means the answer names you with a material error, such as the wrong location, service, audience, or availability.

Add competitor names and the claims attached to each company. This is often the most useful column: knowing exactly which competitor wins the answer for a given question tells you whose pages to study and what gap to close. A competitor appearing more often isn't automatically a problem if the prompt fits its offer better. The useful comparison is whether your company is present and accurately represented for the jobs you want.

Save screenshots when the interface displays citations or search context, because those elements can change after the answer is generated. Add a short confidence note when the conditions were unusual: a logged-in personal account, browsing that was unavailable, or an answer that changed when repeated. Keep the raw record alongside your interpretation, so you can distinguish what the model said from what you think caused it.

Check ChatGPT against Gemini, Claude, Copilot and Perplexity

Your brand can be recommended in one assistant and absent from another, so check the ones your buyers actually use rather than ChatGPT alone. A sensible starting set is ChatGPT, Gemini, Perplexity and Google's AI Overviews, plus Microsoft Copilot if your buyers work in Microsoft 365 and Claude or Grok if your audience leans that way. Perplexity is worth including for another reason: it shows its citations, so you see which source it trusted, not only whether you appeared.

Hold the buyer prompt and scoring rules constant, then treat each platform as a separate environment rather than combining all results into one visibility score. Record whether browsing or citations were available, because a response generated without current web access is not directly comparable with one that searched the web. Keep the location and language consistent, and note any platform that does not expose a model or search state.

Compare patterns, not isolated answers. Create one comparison row per prompt showing which assistants named you, which described you correctly, and which gave a usable link. If every assistant omits you, the problem is broad discoverability or unclear positioning. If one names you accurately and the others don't, the result probably depends on which sources each one reaches. If they mention you but disagree about what you do, your public information is likely inconsistent.

Official product documentation is the right place to check how search, citations, and model selection currently work. Those features change. Your tracking sheet should therefore include the platform configuration for every test, so a later change in results can be linked to a changed environment instead of blamed immediately on your website.

Is the answer based on your website?

You cannot prove that an answer used your website merely because it sounds familiar, so check the cited evidence and compare specific facts. Look for a source link, quoted wording, service description, location, or policy you can trace to a public page, then open that page and confirm it actually supports the claim.

Ask a follow-up evidence question, such as “Which page confirms this service?”, only after you have saved the original answer, and treat the reply as another observation rather than proof. Mark each material claim as supported by an accessible page, plausible but unverified, or contradicted by your public information.

The practical gotcha is stale information. An assistant may repeat an old address, a discontinued service, or a third party's description of you. When the answer describes your company wrongly, or confuses you with a business of a similar name, that is an identity problem with its own fixes, covered in how AI sees your company and how to correct a wrong answer.

Tell a visibility problem from an answer problem

Separate retrieval failure from representation failure by checking whether the assistant can find your business and whether it describes the business correctly. These are different problems and need different evidence.

If your brand never appears for a relevant prompt, check whether independent pages describe your category, service area, audience, and use cases clearly. If your brand appears but the details are wrong, the tracking result is a representation problem. Record the exact false claim and the page or profile that should correct it.

If ChatGPT cites your website but ignores the relevant service page, the problem may be source selection or page clarity rather than brand absence. If it recommends you only after you name the company, the test shows recognition after prompting, not organic discovery. Label that result separately.

The common failure mode is treating every weak answer as a content task. Sometimes the prompt is outside your actual offer, the location is ambiguous, or the model has no reliable source to consult. Check the prompt fit, search context, and cited evidence before changing a page. A clean diagnosis prevents random edits based on one surprising response.

When should you trust an AI visibility score?

Trust an AI visibility score only when you can see the prompts, platform conditions, scoring rules, and underlying answers behind it. A single percentage or rank hides the difference between a passing mention and an accurate recommendation.

A useful score is an internal trend indicator. For example, you might assign separate points for being mentioned, recommended, accurately described, and supported by a relevant citation. Keep the rules fixed, document them in the sheet, and report the component results beside the combined score.

Do not compare scores from two tools unless they test the same platforms, prompts, locations, model settings, and definitions. One tool may count any mention, while another may count only an uncued recommendation. Neither score is automatically wrong, but they answer different questions.

Treat results as directional because model outputs are probabilistic and interfaces change. A score that rises after one run may reflect prompt variation or a temporary source choice. Look for a pattern across repeated checks and inspect the raw answers before taking action. The score is a filing label for evidence, not proof that a specific number of buyers will find you.

How often to check, and when to automate

Check on a fixed cadence you can keep, monthly at least and weekly if your category moves fast, and again after any material change: a new service, a new location, a rebrand, or a rewrite of an important page. A single check is a snapshot of a moving target. Answers change as models update, as competitors sharpen their pages and as new sources get cited, so the trend line across checks is what tells you whether your work is landing.

The manual routine is the right way to learn what to watch, but it grows quickly. Ten prompts across six assistants is sixty answers to run, read and log every cycle. That is the point to consider a monitoring tool, which runs the same grid on a schedule and shows only what changed. Tools in this category include Cituna, Otterly.AI and Peec AI. Cituna goes furthest past tracking: it checks buyer prompts on seven engines every day, connects to Google Search Console so an answer can be read beside the clicks behind it, writes the fix for each gap, and turns the sites the engines cite into an outreach plan. Whether you need one yet is weighed in whether to pay for an AI visibility tracking tool, and what each one charges is in what an AI visibility platform costs.

What to do with the results this week

Choose one corrective action from the most repeatable failure, then rerun the same prompt panel after the change. If the model cannot identify what you do, clarify the category, audience, locations, and services on your core pages. If it gives outdated facts, correct those facts at the source and on relevant business profiles.

If your brand is mentioned but not recommended, compare the reasons given for competing providers. Look for a missing, supportable detail such as a clear service boundary, suitable customer type, process explanation, or proof of a specific capability. Do not add claims merely because a model seems to prefer them. The fixes that tend to change these answers are set out in how to improve your brand's visibility in AI.

If citations point to irrelevant pages, improve the page that should answer the buyer’s question and make its title, headings, and supporting details unambiguous. If the answer repeatedly invents information, log the error and avoid treating the model as a dependable source for that claim.

Rerun the unchanged prompts and record the difference. Keep the old answers. A before-and-after comparison is more useful than a new result with no baseline. The weekly habit is simple: test, preserve evidence, diagnose one failure, make one justified change, and test again.

Sources consulted

• OpenAI Help Center (help.openai.com)
• Anthropic documentation (docs.anthropic.com)
• Microsoft Support (support.microsoft.com)
• Cituna pricing (cituna.com)

Compare AI tools on price