How to measure brand visibility in ChatGPT and AI answers
You measure brand visibility in ChatGPT by asking it what your buyers ask, the same way every time, and writing down what comes back. Build a fixed panel of buyer prompts, run each one several times, and score every answer on three levels: whether your brand is mentioned, whether it is recommended, and whether one of your pages is cited as a source. Count your mentions against those of named rivals to get your share of answer. Then add the free first-party data: Search Console's generative AI report for Google's AI Overviews and AI Mode, the AI Performance report in Bing Webmaster Tools for Copilot and Bing, ChatGPT and Perplexity referrals in your analytics, and AI agents' visits in your server logs. None of it requires buying software.
Book a 30-minute call if you would like to see your starting point before you build the sheet: we will go through your category with you and show what the engines answer today when a buyer asks for a recommendation.
Mentioned, recommended, cited
"Visibility" in an AI answer is three different things, and each points to a different fix, so score them separately.
| Level | What counts | What a gap usually points to |
|---|---|---|
| Mentioned | Your brand is named anywhere in the answer | The engine does not yet connect you with the category |
| Recommended | You are put forward as a fit for the need the buyer stated | It knows you exist, but not why you suit this buyer |
| Cited | A page on your domain is among the answer's linked sources | Your own pages are not part of the evidence it used |
Record a fourth thing alongside them: whether what the answer says about you is right. A recommendation built on last year's price or a retired feature has to be fixed where the engine found it.
A brand can rank well in Google and still go unnamed in these answers. AEO vs SEO: what actually changes explains why the two come apart.
Step 1: Build a fixed panel of buyer prompts
Write twenty to fifty prompts the way buyers phrase them: full sentences that carry the constraints a shortlist depends on, such as team size, budget and the tools already in place. Spread them across five groups.
| Group | Example prompt |
|---|---|
| Category | "What's the best [category] software for a 40-person B2B SaaS company?" |
| Alternatives | "What are good alternatives to [rival] for a team that needs [capability]?" |
| Comparison | "[Your brand] or [rival]: which suits [use case] better?" |
| Problem | "How can a [role] team [do the job] without [the pain]?" |
| Brand facts | "How much does [your brand] cost, and what does each plan include?" |
Then pick three to five rivals to score against: the ones your sales team meets in deals, not the ones you admire. Freeze the wording, save the panel with a version number, and treat any reworded prompt as a new one. A trend only means something on an unchanged panel.
Step 2: Run every prompt the same way, more than once
Start with ChatGPT, then add the engines your buyers use: Perplexity, Gemini, Claude, Grok, and Google's AI Overviews and AI Mode. Keep the conditions identical from run to run: a fresh conversation, memory and personalisation switched off where the product allows it, the same mode (with or without web search), the same type of account and, where you can manage it, the same location. Note the date of every run.
Run each prompt at least three times per engine in every cycle. Answers are written afresh each time and the shortlist can change between runs. Bing, describing its own citation data, says the mix of cited sources can shift as models evolve, as freshness signals change and as user behaviour moves (Bing Webmaster Blog, June 2026). One run is an anecdote; three are the start of a rate.
Save the full text of each answer and its list of sources. A screenshot is evidence; a row in a sheet is data.
Step 3: Score every answer in one sheet
One row per answer, with the same columns every cycle:
| Column | What to record |
|---|---|
| Prompt | Its ID and group from the frozen panel |
| Run | Engine, mode, date and run number |
| Mentioned | Yes or no |
| Position | Where you first appear among the options named: first, second, third |
| Recommended | Yes or no: put forward as a fit for the stated need |
| Cited | Yes or no, and which of your URLs |
| Rivals named | Each named rival, in the order they appear |
| Sources | Every domain the answer links to |
| Accuracy | Anything wrong about you: price, plans, category, features |
Keep the sources column even when it feels tedious. Over a few cycles it becomes a map of the pages each engine relies on for your category, and those pages are where you can work to change the answers.
Step 4: Turn the sheet into four numbers
- Mention rate: answers that name you, divided by all answers.
- Recommendation rate: answers that recommend you, divided by all answers.
- Citation rate: answers that cite a page on your domain, divided by the answers that show sources.
- Share of answer: your mentions, divided by all mentions of you and your named rivals.
A made-up example, to show the arithmetic only: the numbers are invented and are not a benchmark. Suppose a panel of 30 prompts run three times in ChatGPT gives 90 answers, all of them with sources. If you are named in 27 of them, your mention rate is 30%; if you are recommended in 18, that is 20%. If a page of yours is cited in 9, the citation rate is 10%. If your three named rivals are mentioned 54, 36 and 18 times across the same answers, your share of answer is 27 of 135 mentions: 20%.
Report each number per engine and per prompt group, never as one blended figure, because an average across engines hides the engine where you are missing. And compare cycle with cycle on the same panel, not with someone else's numbers from a different panel.
Step 5: Add the free first-party signals
Some of this the platforms report themselves, at no cost. Each signal covers only its own surface, so it adds to the panel rather than replacing it. The platform documentation cited below was read on 16 October 2026.
Google Search Console. Google introduced its generative AI performance report on 3 June 2026 and had rolled it out to all websites worldwide by 31 August 2026 (Search Central blog). It counts impressions in AI Overviews and AI Mode, which Google defines as the times a link to your site was shown in one of those features, and splits them by page, country, device and date. Several links from your site in one feature count once in the chart total, and the report documentation describes impressions only. So it tells you where Google showed you, not what the answer said about you. If the report is missing, the documentation gives the usual reason as too few impressions in those features. Discover has a report of its own.
Bing Webmaster Tools. Bing's AI Performance report went into public preview on 10 February 2026 (Bing Webmaster Blog). It shows how often your pages are cited in Microsoft Copilot, in Bing's AI-generated summaries and in some partner integrations: total citations, the average number of your pages cited per day, a sample of the grounding queries behind those citations, and citation counts for each URL. In June 2026 Bing added Intents, Topics, Citation Share and a Compare view, all in preview. Citation Share is your percentage of all the citations shown for a grounding query; Bing does not say who holds the rest.
Analytics referrals. Build a segment for sessions referred by chatgpt.com or perplexity.ai, and include landing URLs tagged utm_source=chatgpt.com, the parameter OpenAI's publisher FAQ stated that ChatGPT adds to the links it sends. That article now redirects to OpenAI's notice retiring ChatGPT Atlas (checked on 16 October 2026), so confirm the tag in your own data before you build reports on it. Some assistant traffic arrives with no referrer at all, from apps or copied links, and is counted as direct, so treat the segment as a floor. The landing pages it reaches show which of your pages the answers send people to.
Server logs. OpenAI and Perplexity both publish the names of their agents and the IP ranges they use. OAI-SearchBot is the crawler behind ChatGPT's search features, and ChatGPT-User can fetch a page while ChatGPT is answering someone's question (OpenAI's crawler documentation). PerplexityBot crawls for Perplexity's search results, and Perplexity-User fetches pages for a user's question (Perplexity's crawler documentation). Visits from the two user-triggered agents to your buying pages are as close as logs get to watching live questions reach you. Check the source addresses against the published lists before trusting a user-agent string. What each Perplexity agent does with robots.txt is covered in how to get cited by Perplexity.
| Signal | What it covers | What it counts | Shows what the answer said? |
|---|---|---|---|
| Your prompt panel | Every engine you run it in | Mentions, recommendations, citations, share of answer | Yes |
| Search Console generative AI report | Google AI Overviews and AI Mode | Impressions by page, country, device and date | No |
| Bing Webmaster Tools AI Performance | Copilot, Bing's AI summaries, some partners | Citations, cited pages, grounding queries, citation share | No |
| Analytics referrals | Visits that arrive with a referrer or tag | Sessions and landing pages | No |
| Server logs | Agents that identify themselves | Fetches, page by page | No |
How often to measure, and how to read the result
Run the panel once a month and look at the first-party signals weekly. Keep the panel unchanged for at least a quarter; when new prompts are needed, add them as a separate group so the original trend still compares like with like. Rerun early after a change that ought to move answers, such as new pricing, a new comparison page or a rebrand.
Then read the gaps:
- Rarely mentioned. The engines do not connect you with the category. Start with the sources column: which pages do the answers cite, and are you on them?
- Mentioned but rarely recommended. The engines know the name but not the fit. Clearer positioning, and pages that say plainly who you are for, usually come first.
- Recommended but rarely cited. Other people's pages are making your case. That is valuable, and it leaves you depending on pages you do not control.
- Wrong facts. Correct them on your own pages first, then on the third-party pages the answers cite.
What to change once you have a baseline is the subject of how to get a SaaS product recommended by AI.
Commercial tracking tools exist that automate these runs across engines. The method underneath is the same, so whichever route you take, make sure you can see the prompts and the raw answers behind every number.
Having it run for you
Measurement is one of three workstreams at The Rank Master, alongside revenue pages and earned citations. It reports in a live dashboard and a monthly report covering the seven engines buyers ask first, ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews and Google AI Mode: who gets named for your buying questions and which sources they quote. Search Console data is joined to it, so clicks and citations can be read side by side. The checkpoints are agreed up front: in month 1 the baseline is taken across all seven engines, by month 3 answer share against named rivals is charted, and by month 6 it is reported engine by engine, with a citation map of the pages the engines quote for your category.
Plans have published pricing, a 90-day initial term, then month-to-month on 30 days' notice, and everything made for you stays yours: pages, briefs, data and contacts leave with you. Book a 30-minute call and we will look at your category live, show you what the engines say when a buyer asks for a recommendation, and tell you straight whether we can help.