AI Visibility Tracking: Monitor Your Brand in ChatGPT and Gemini

In short: AI visibility tracking means asking ChatGPT, Gemini, Perplexity and Claude the same fixed set of customer questions on a schedule and measuring how often they mention you, cite your site, rank you in their recommendations and describe you positively, compared with competitors. Because answers change from run to run, you track rates over many answers, not a single result. To improve, make your site readable by AI search crawlers, state clearly who you are, and earn mentions on the third-party pages assistants cite.

When someone asks ChatGPT "what's the best bookkeeping app for freelancers?", the answer usually names a handful of products. If yours is not one of them, you lost that customer before they ever saw a search result. There is no impressions report for this, no Search Console for chat answers. The only way to know is to ask, systematically, and count.

This guide shows how to do that properly: what to measure, how to build a prompt set, the formulas we use, how to tell a real change from noise, and what actually moves the numbers.

What AI visibility means

"AI visibility" is not one number. It is five separate signals, and they fail for different reasons.

A brand can be mentioned often but never cited (the model knows you from training data, but its live web search pulls other pages). Another can be cited constantly as a source for definitions while never being recommended as a product. Treat each signal separately or you will fix the wrong thing.

Why answers vary, and why one check proves nothing

If you ask an assistant the same question twice, you will often get two different lists. That is expected, for several reasons:

1. Sampling. Language models generate text by sampling from probabilities. Unless the provider fixes every setting, wording and the order of recommendations shift between runs.
2. Retrieval. With web search on, the assistant fetches pages at answer time. Which pages it finds depends on the search results that day, so the sources, and therefore the brands, change.
3. Search on or off. The same model answers from training data alone when it does not search. Those answers lag the web by months and often favor older, larger brands.
4. Model version. Providers update models regularly. A new version can reshuffle recommendations overnight.
5. Location and language. Assistants with web search adapt to the user's country and language. "Best payroll software" in the UK and in the US are different markets.
6. Personalization. ChatGPT can reference saved memories and past chats to tailor replies when the user has memory turned on. Your own logged-in answer is not what a new prospect sees.

The practical consequence: a screenshot of one answer is an anecdote. To measure visibility you need a fixed prompt set, asked repeatedly, from controlled settings (country, language, search on), and you report rates over many answers.

Step 1: Build a prompt set

Your prompt set is your keyword list for AI search. Write it the way customers ask, in full sentences, and leave your brand name out of everything except the brand prompts.

Four prompt types

• Category prompts ("What are the best bookkeeping apps for freelancers?"). These decide whether you are in the consideration set. Weight them most.
• Problem prompts ("How do I track expenses as a freelancer without an accountant?"). The user has not chosen a product category yet. Mentions here are rarer and more valuable.
• Comparison prompts ("Ledgerly vs QuickBooks for a one-person business" or "alternatives to [competitor]"). These show how you are framed against rivals.
• Brand prompts ("Is Ledgerly good?", "What does Ledgerly cost?"). These test accuracy and sentiment, not visibility. Keep them in a separate group so they do not inflate your mention rate, since the brand is named in the question.

Cover the buyer journey

Map prompts to stages: problem awareness, solution research, shortlist, decision. A healthy set for one product in one market has roughly 40% category, 25% problem, 20% comparison and 15% brand prompts. Start with 20 to 40 prompts; you can grow later, but every prompt you add multiplies cost.

Countries and languages

If you sell in several markets, do not translate your English prompts word for word. Write them the way people in that market ask, and run them with that country and language. Report each market separately; a global average hides the fact that you might be strong in the US and absent in Germany.

Where to find real prompts

• Sales and support calls: the exact words prospects use.
• Your top organic keywords, rewritten as questions.
• "People also ask" boxes and community threads in your category.
• The comparison terms competitors target on their own pages.

Step 2: Choose the metrics and formulas

Run every prompt on every assistant you track, several times per period. Then calculate:

• Mention rate = answers that name your brand ÷ total answers (excluding brand prompts).
• Citation rate = answers that link to your domain ÷ total answers.
• Average position = mean of your rank in the recommendation list, counted only in answers that contain a list and mention you. Report it together with mention rate, because a 1.5 average position on 5% of answers is not a win.
• Top-3 rate = answers where you are in the first three recommendations ÷ total answers. This combines presence and position in one number and is often the most useful headline metric.
• Share of voice = your mentions ÷ (your mentions + mentions of every tracked competitor).
• Sentiment = share of your mentions classified as positive, neutral or negative. Read a sample by hand each month to check the classifier.

Also log every cited URL, yours and others'. The list of domains the assistants cite instead of you is your outreach and content plan.

A worked example (hypothetical numbers)

The following is a hypothetical example, not real data. Ledgerly, an invented bookkeeping app for freelancers, tracks 20 non-brand prompts on four assistants in the US, three runs each per month: 20 × 4 × 3 = 240 answers.

Two invented competitors appear in the same answers: Competitor A in 132 answers and Competitor B in 96. Share of voice:

• Ledgerly: 84 ÷ (84 + 132 + 96) = 84 ÷ 312 = 26.9%
• Competitor A: 132 ÷ 312 = 42.3%
• Competitor B: 96 ÷ 312 = 30.8%

What this tells Ledgerly: Perplexity, which is built around live web search, cites it most, so its pages are retrievable. Claude and Gemini mention it least, which suggests weaker presence in the third-party sources those answers draw on. The next step is to open the answers where Competitor A appears and Ledgerly does not, and list the URLs they cite.

How big a change is real?

Rates from a sample carry noise. A quick rule of thumb uses the standard error of a proportion, √(p × (1 − p) ÷ n):

• At 35% across 240 answers, the standard error is about 3 points, so a month-to-month swing of ±6 points can be pure chance.
• At 40% on 60 ChatGPT answers, the standard error is about 6 points, so per-assistant swings of ±12 points mean little on their own.

Decision rule: act on a change that is larger than twice the standard error, or that persists in the same direction for three periods. Do not rewrite your homepage because one week dipped.

Step 3: Collect the data

The manual method

A spreadsheet works for a first audit:

1. Create columns for date, assistant, country, prompt, prompt type, run number, mentioned (yes/no), position, cited (yes/no), cited URLs, competitors named and sentiment.
2. Use a clean browser profile, logged out or with memory off, and set the assistant to search the web.
3. Paste each prompt, record the answer, repeat.

It is free and teaches you a lot about how the assistants reason. It also stops scaling quickly: 20 prompts, four assistants and three runs is 240 copy-and-paste rounds per month, and consumer apps make it hard to control location.

Using a tool

AI visibility tools send the prompts through the providers' interfaces on a schedule, parse the answers and calculate the metrics. When you evaluate one, check:

• Which assistants it covers and whether web search is on.
• Whether it controls country and language per prompt.
• Whether it stores the full answer and cited URLs, not just a score.
• How it handles brand name variants and misspellings.
• Whether it tracks competitors' mentions in the same answers.

Note that answers collected through an API with web search on are a close proxy, not a copy, of what a logged-in user sees in the consumer app. That is fine for trends, which is what you are measuring.

Step 4: Improve your AI visibility

Nobody outside the providers knows exactly how each assistant chooses brands. What we can do is follow what the providers document, and observe what the cited sources have in common.

Let the right crawlers in

OpenAI separates its crawlers by purpose. Per OpenAI's crawler documentation:

• OAI-SearchBot surfaces websites in ChatGPT's search features. Sites that opt out of it "will not be shown in ChatGPT search answers."
• GPTBot crawls content that may be used to train OpenAI's foundation models. Disallowing it opts you out of training.
• ChatGPT-User fetches a page when a user's request needs it; it is not an automatic crawler, and robots.txt rules may not apply.

The settings are independent, so a common choice is to block GPTBot and allow OAI-SearchBot. OpenAI notes changes can take about 24 hours to take effect.

Perplexity's documentation describes PerplexityBot, which surfaces and links websites in Perplexity search results and is not used to train foundation models, and Perplexity-User, which fetches pages for a user's question and generally ignores robots.txt. Perplexity recommends allowing PerplexityBot if you want to appear in its results.

For Google, AI features and your website is explicit: there are no additional requirements or special optimizations to appear in AI Overviews or AI Mode, the usual SEO practices apply, and you do not need new machine-readable files, AI text files or markup. Traffic from these features is reported in Search Console under the "Web" search type. Separately, the Google-Extended token controls whether content Google crawls may be used to train future Gemini models and for grounding in Gemini Apps and Grounding with Google Search on Vertex AI; Google states it does not affect inclusion or ranking in Google Search.

The most common failure we see is not robots.txt at all. It is a CDN or firewall rule that silently blocks AI user agents while robots.txt says "allow." Check both. Our free AI crawler checker reads robots.txt for the major AI crawlers and sends a test request with ChatGPT's crawler name to see what your firewall returns.

Make your entity unambiguous

Assistants need to know what you are before they can recommend you. On your own site:

• Say in one plain sentence, near the top of the homepage, what the product is, who it is for and where it is available.
• Keep pricing, plans and feature pages current and crawlable as HTML, not only inside JavaScript or images.
• Add Organization structured data. Google says it helps distinguish your organization from others with similar names, and its sameAs property links your site to your profiles elsewhere (Organization structured data). The schema markup checker shows what your page currently declares.
• Publish clear, quotable pages for the questions in your prompt set: comparisons, "how to" guides, specific use cases.

Be present where the assistants look

When assistants search the web, they cite review sites, comparison articles, community threads, documentation and news. In our experience, the brands with high mention rates are the ones that show up consistently across those third-party pages, not the ones with the most content on their own site. Practical steps:

1. Export every URL cited in answers where a competitor appears and you do not.
2. Group them: review platforms, listicles, forums, directories, publications.
3. For listicles, contact the author with accurate, current information. For review platforms, ask real customers to review you. For communities, answer questions honestly where you have expertise, following each community's rules.
4. Fix outdated or wrong information about you wherever it appears. Assistants repeat what the web says.

What about llms.txt?

llms.txt is a proposal, first published by Jeremy Howard in September 2024, for a Markdown file that gives language models a short guide to a site: an H1 title, a one-line summary and lists of links to key pages. It is easy to create and harmless. Be realistic, though: Google says AI text files are not needed for its AI features, and as of October 2026 we have not found a public commitment from OpenAI, Anthropic or Perplexity to use llms.txt when answering. Add one if it takes you ten minutes, then spend your real effort elsewhere. The llms.txt checker validates the format.

Common mistakes

• Putting your brand in category prompts. "Is Ledgerly the best bookkeeping app?" will mention Ledgerly every time and tell you nothing.
• Measuring once. A single run is a sample of one. Track at least monthly, ideally weekly, with repeated runs.
• Averaging across countries. Report each market separately.
• Chasing every dip. Use the noise rule above before you change anything.
• Blocking the wrong bot. Blocking OAI-SearchBot to avoid training removes you from ChatGPT search answers; GPTBot is the training crawler.
• Ignoring sentiment. Being mentioned as "cheaper but buggy" is visibility you do not want. Fix the underlying issue, then the sources.

A 30-minute starting checklist

1. Write 20 non-brand prompts across category, problem and comparison types, plus five brand prompts.
2. Pick three competitors you lose deals to.
3. Run the free AI visibility checker for a first read on ChatGPT and Perplexity.
4. Run the AI search readiness checker to see whether crawlers can read your homepage and whether your identity markup is in place.
5. Record mention rate, citation rate, top-3 rate and share of voice as your baseline.
6. List the top ten domains cited instead of you. That is your first month of work.

How Searvia tracks AI visibility

Searvia's AI visibility tracking asks your questions every week with web search turned on, on ChatGPT, Gemini, Perplexity and Claude on paid plans (Perplexity on the free plan). For each answer it records whether you are mentioned, recommended or cited, your position in the recommended list, the tone of the description and your share of voice against competitors, and it lists the sites cited instead of you. It also checks whether AI bots can reach your site through robots.txt, firewall rules and llms.txt. The AI visibility help guide walks through setup in about four minutes, and you can start with the free checker before creating an account.