How to Measure Brand Visibility in ChatGPT and AI Search: Method and Template
The method Ranking Atlas uses for every baseline: prompt design, engine runs, a log you can query, source classification and a one-page monthly report.
On this page (14 sections)
Brand visibility in AI search is the share of buyer questions on which ChatGPT, Gemini, Perplexity and Google's AI answers name your brand. You measure it the way you would measure any sample: a fixed panel of questions, repeated runs, a log, and a baseline to compare against.
This page is the full method, with the log template. For a first read on your own brand, a call gives you a live review of how AI search sees your website today.
How Do You Measure Brand Visibility in ChatGPT and AI Search?
Measure brand visibility in AI search in five steps: build a fixed set of buyer prompts, run them across the engines on a schedule, log every answer in a structure you can query, classify each cited source, then record a baseline and report the trend against it. Keep branded prompts separate throughout, so the score cannot hide a gap.
- Build the prompt set: about fifty buyer questions, branded prompts kept apart from non-branded ones, frozen once built.
- Run the engines: every prompt on ChatGPT, Perplexity, Gemini and Google AI Overviews, from a clean context, at least twice per run, on a fixed cadence.
- Log the answers: one row per prompt, engine and run, recording whether you are named, where, what is said, and which URLs are cited.
- Classify cited sources: earned, owned or rented; recurring or one-off; durable or exposed.
- Baseline, then trend: write the first run down before any work starts, then report movement against it month by month.
How Do You Check Your Brand's Visibility in ChatGPT?
To check your brand's visibility in ChatGPT, turn search on, start each run from a clean context with no history or personalisation, set the location to your market, and ask a fixed set of buyer prompts. Log whether your brand is named, which URLs are cited and which competitors appear, then repeat the same prompts on a schedule.
-
Turn search on
Answers built from web search show the sources behind them, which is half of what you are measuring.
-
Start from a clean context
No history, memory or personalisation, so your own usage does not shape the answer. A fresh temporary chat for each prompt.
-
Set the location to your market
A question asked from the UK retrieves different sources from the same question asked from the US. Record the country in the log.
-
Ask the fixed buyer prompts
The same questions every time, written the way a buyer types them, at least twice each.
-
Log named, cited and competitors
Whether your brand is named, which URLs are cited, and who is named alongside you.
The same steps apply to Perplexity, Gemini and Google AI Overviews. Only the interface changes.
What Metrics Matter Most When Measuring Brand Visibility Across AI Platforms?
Five metrics carry an AI visibility report: non-branded visibility rate (the share of buyer-question answers that name you), branded rate as a control, share of voice against each named competitor, citation rate (the share of answers that cite a page you own), and source concentration (which domains decide the answers). Everything else is decoration.
| Metric | Formula | What it tells you |
|---|---|---|
| Non-branded visibility rate | Answers to non-branded prompts that name you ÷ all non-branded answers, per engine | Whether buyers who have not heard of you get your name |
| Branded rate (control) | Answers to branded prompts that name you ÷ all branded answers | Whether the engines know who you are at all; usually high, reported separately |
| Share of voice | Non-branded answers naming each brand ÷ all non-branded answers, for you and each competitor | Who the engines recommend instead of you, and by how much |
| Citation rate | Answers citing a URL you own ÷ all answers | Whether your own pages are retrieved, as distinct from your name being known |
| Source concentration | Share of all citations taken by the ten most-cited domains, per engine | How few sources decide your category, and therefore where the work goes |
- Formula:
- Answers to non-branded prompts that name you ÷ all non-branded answers, per engine
- What it tells you:
- Whether buyers who have not heard of you get your name
- Formula:
- Answers to branded prompts that name you ÷ all branded answers
- What it tells you:
- Whether the engines know who you are at all; usually high, reported separately
- Formula:
- Non-branded answers naming each brand ÷ all non-branded answers, for you and each competitor
- What it tells you:
- Who the engines recommend instead of you, and by how much
- Formula:
- Answers citing a URL you own ÷ all answers
- What it tells you:
- Whether your own pages are retrieved, as distinct from your name being known
- Formula:
- Share of all citations taken by the ten most-cited domains, per engine
- What it tells you:
- How few sources decide your category, and therefore where the work goes
Report every rate with its denominator, engine and date. A rate from fewer than 30 answers is directional.
How Should Branded and Non-Branded Prompts Be Separated in AI Visibility Reporting?
Branded and non-branded prompts should be tagged at setup and reported as two separate lines, never averaged. A prompt containing your name retrieves your own site and coverage about you, so you appear almost automatically. A buyer prompt ("best [category] for [segment]") retrieves the sources the engines trust on the category, and appearing there has to be earned.
A brand can have strong visibility on branded prompts and close to none on buyer questions. A blended score averages those into a comfortable middle number that measures nothing.
The split is the load-bearing wall of the method, which is why it appears in every step below. Tag each prompt once, at setup, and the tag follows every row of the log.
How Do You Build the Prompt Set?
Build the prompt set from about fifty buyer questions in three blocks: twenty non-branded commercial prompts, twenty non-branded informational prompts, and ten branded and competitor prompts kept as a control. Source the wording from sales calls, support tickets, Search Console and win-loss notes, written the way a buyer types, then freeze the set.
| Block | Roughly | Examples | Why it is there |
|---|---|---|---|
| Non-branded commercial | 20 | "best [category] for [vertical]", "which [category] providers work with [buyer type]" | The shortlist-forming questions. Buyers phrase them as questions to a knowledgeable colleague, not as keywords. |
| Non-branded informational | 20 | "how do [buyer type] solve [problem]", "how to evaluate [category]" | Research-phase questions that show which sources own the education layer of your category. |
| Branded and competitor | 10 | your name, competitors' names, direct comparisons | A control group, reported separately and never blended. |
- Roughly:
- 20
- Examples:
- "best [category] for [vertical]", "which [category] providers work with [buyer type]"
- Why it is there:
- The shortlist-forming questions. Buyers phrase them as questions to a knowledgeable colleague, not as keywords.
- Roughly:
- 20
- Examples:
- "how do [buyer type] solve [problem]", "how to evaluate [category]"
- Why it is there:
- Research-phase questions that show which sources own the education layer of your category.
- Roughly:
- 10
- Examples:
- your name, competitors' names, direct comparisons
- Why it is there:
- A control group, reported separately and never blended.
Search Console is the best source of real wording. Long questions typed into Google ("how do I know how visible we are in ChatGPT") are the same questions typed into ChatGPT. The 34 questions behind the figures on this page came from there.
Which Engines Should You Run, and How?
Run every prompt on ChatGPT with search on, Perplexity, Gemini and Google AI Overviews, because they retrieve from different source sets and a brand's presence varies across them more than most teams expect. Run from a clean context, from your buyers' location, at least twice per prompt per run, monthly at minimum.
How different are the engines? Ranking Atlas, a research-grade AI SEO agency, ran 34 buyer questions twice each on ChatGPT, Gemini and Perplexity on 5 October 2026. The 204 answers cited 1,978 sources on 762 distinct domains.
Of 762 cited domains, 652 (86%) were cited by one engine only, 90 (12%) by two engines and 20 (3%) by all three.
Source: Ranking Atlas sample, 5 October 2026: 34 buyer questions, two runs each, ChatGPT (gpt-5.5, web search), Gemini (2.5 Flash, Google Search grounding) and Perplexity (sonar), via the DataForSEO LLM Responses API, US location.
Twenty domains in 762 were cited by all three engines. A brand present in ChatGPT's sources can be absent from Gemini's, which is why one engine never stands in for the others.
The engines also differ in how much they cite. Gemini returned a median of 13 sources per answer, ChatGPT 6, and six of ChatGPT's 68 answers cited nothing. Perplexity returned ten every time, which looks like a limit of the API route rather than the engine.
Run each prompt at least twice. In the same sample, only 55% of the domains cited in a first run were cited again in a second run minutes later: 29% on ChatGPT, 42% on Gemini, 92% on Perplexity. The full run-to-run table is in our comparison of AI visibility tools.
How Do You Log the Answers so the Data Can Answer Questions?
Log one row per prompt, engine and run, in a flat table with fixed columns: date, prompt ID, prompt type, engine, country, named (yes or no), position, what was said, competitors named, cited URLs and source class. Every metric in this guide then falls out with a filter and a count, in a spreadsheet or a database.
| Column | Values | Example row |
|---|---|---|
| date | ISO date of the run | 2026-10-05 |
| prompt_id | Stable ID; versions of the set get a suffix | NBC-07 (v1) |
| prompt_type | branded, non-branded commercial, non-branded informational | non-branded commercial |
| prompt_text | Verbatim | "best endpoint security platform for a 500-person company" |
| engine | chatgpt, perplexity, gemini, aio | chatgpt |
| run | 1, 2, 3 within the same session | 2 |
| country | Location the run was made from | GB |
| named | yes or no | no |
| position | first, listed, absent | absent |
| said | One line on how the brand was described, if named | |
| competitors_named | Brands named, in order | Vendor A; Vendor B; Vendor C |
| cited_urls | Every URL cited, in order | 8 URLs |
| source_class | Per URL: earned, owned or rented; recurring or one-off | earned/recurring; rented/one-off; … |
| aio_shown | AI Overviews only: whether an Overview appeared | n/a |
Source: Ranking Atlas template. The example row uses made-up values for a cybersecurity vendor.
Two habits keep the log honest. Paste the full answer text into a second sheet, keyed by date, prompt ID, engine and run, so any figure can be traced back to what the engine said. And never edit a past row; add a new run.
How Do You Classify Every Cited Source?
Classify every URL the engines cite three ways: earned, owned or rented (who controls it); recurring or one-off (whether the engines return to it); and durable or exposed (whether it survives platform enforcement). Then build the source map: the domains that decide the answers in your category, and which of them name you.
| Question | The split | What it tells you |
|---|---|---|
| Earned, owned or rented? | Earned: editorial coverage, genuine community discussion, independent references. Owned: your domain and properties. Rented: paid placements, self-published rankings positioned as neutral, seeded community content. | Which citations nobody could buy, and which exist because someone paid for or manufactured them. |
| Recurring or one-off? | Whether the domain is cited across several prompts and runs, or appeared once. | A citation that appears once tells you less than one the engines return to constantly. |
| Durable or exposed? | Earned editorial survives enforcement waves; rented sources are actively hunted, and community platforms delete manufactured content retroactively. | Which parts of your visibility, and your competitors', are load-bearing, and which a platform action removes overnight. |
- The split:
- Earned: editorial coverage, genuine community discussion, independent references. Owned: your domain and properties. Rented: paid placements, self-published rankings positioned as neutral, seeded community content.
- What it tells you:
- Which citations nobody could buy, and which exist because someone paid for or manufactured them.
- The split:
- Whether the domain is cited across several prompts and runs, or appeared once.
- What it tells you:
- A citation that appears once tells you less than one the engines return to constantly.
- The split:
- Earned editorial survives enforcement waves; rented sources are actively hunted, and community platforms delete manufactured content retroactively.
- What it tells you:
- Which parts of your visibility, and your competitors', are load-bearing, and which a platform action removes overnight.
The recurring-or-one-off split matters more than it looks. In our 5 October 2026 sample, 405 of the 762 cited domains (53%) were cited exactly once across 204 answers.
Domains cited in only one answer: 70% of ChatGPT's 221 cited domains, 61% of Gemini's 472, and 16% of Perplexity's 199.
Source: Same sample as Figure 1. One-off domains: ChatGPT 155 of 221 cited domains, Gemini 286 of 472, Perplexity 32 of 199.
The ten most-cited domains took 22% of ChatGPT's citations, 23% of Perplexity's and 12% of Gemini's. The source map for a category is short at the top and very long at the bottom, and the work goes to the top.
Check the map against the marketing plan. The sources deciding a category are not always the ones the plan targets, and LinkedIn and Reddit were both among the twenty most-cited domains in this sample.
How Do You Benchmark Brand Visibility Against Competitors in ChatGPT?
Benchmark against competitors by running the same fixed prompt set, logging every brand named in every answer, and calculating share of voice: the proportion of non-branded answers that name each brand, per engine. Compare the shares against your baseline month by month, never on a single run, because answers vary between runs.
A worked example, with made-up figures. A cybersecurity vendor runs 40 non-branded prompts twice on ChatGPT: 80 answers.
- Vendor A is named in 36 answers: share of voice 45%.
- Vendor B is named in 24: 30%.
- You are named in 8: 10%.
- Branded control: on the 10 branded prompts you are named in 19 of 20 answers (95%), which is why that line is reported separately.
Shares do not sum to 100%, because one answer can name several brands. The gap that matters is 45% against 10% on the questions buyers ask before they know any of you.
Three rules keep the comparison fair. Same prompts, same engines, same month for every brand. Count a brand as named only when the answer text names it, not when a page about it is cited. And state the denominator every time: "8 of 80 answers", not "10%".
How Do You Track Brand Presence in Google AI Overviews?
Track Google AI Overviews with the same prompt set and log, plus one extra column: whether an Overview appeared at all for that prompt. Overview coverage expands and contracts, and being cited in an Overview that stopped appearing is a different event from losing the citation. Report presence as a share of prompts that showed an Overview.
Run from your buyers' country, logged out, and record the result of each prompt even when no Overview shows. Google AI Mode is a separate surface; if your buyers use it, log it as its own engine.
What Evidence Belongs in an AI Visibility Report?
An AI visibility report needs six things: non-branded visibility rate this month against baseline, per engine; branded rate alongside it, labelled as the control; share of voice against each named competitor on the same prompts; new sources citing you, classified; sources you lost; and one paragraph each of interpretation and next month's targets. One page.
- Non-branded visibility rate this month versus baseline, per engine, with the denominator.
- Branded rate, labelled as the control, so nobody reads it as progress.
- Share of voice against each named competitor on the same prompt set.
- Sources gained: new domains citing you, classified earned, owned or rented.
- Sources lost, with the prompts they used to appear on.
- Interpretation: one paragraph, including luck in both directions, and one paragraph of what next month targets.
Any report longer than that is usually padding; anything missing from it is usually a problem. Which metrics belong in the report and which decorate it is the subject of our measurement KPI guide.
Why Record a Baseline Before Any Work Starts?
Record a baseline because success defined after the fact drifts towards whatever happened. Run the full set once before any visibility work begins and write down non-branded rate per engine, branded rate, share of voice per competitor and the source map. Every later report is a comparison against that document, not a story.
Then re-run monthly and evaluate at programme level. Answers vary run to run by design, engines change behaviour without notice, and single-month movements are weather. The trend across months is climate.
Visibility compounds as each round of coverage extends the citation base the engines retrieve from, so the measurement window has to match the mechanism: quarters, not weeks.
What Can This Method Not Do?
This method cannot make a single run authoritative, because generative answers are probabilistic; only the panel and the trend are evidence. It cannot attribute revenue to individual citations with precision. And it does not move the numbers. It makes them honest, and it shows where the work that moves them should go.
Closing a non-branded gap means building pages the engines can retrieve and earning corroborating citations in the sources the map surfaces. That work is covered in our guides to how to get cited by ChatGPT and citation equity, and at the mechanics level in our GEO versus SEO analysis.
Measurement without the building is a well-documented flatline. Building without the measurement is spend without evidence. Ranking Atlas pairs them in its AI SEO programme for B2B software companies, baseline first, and publishes the method in full because a buyer who measures properly, with anyone, makes the whole discipline more honest. If you would rather have it done, our guide to AI SEO agencies for B2B software companies sets out how to choose.
FAQ
How often should the prompt set run?
Monthly at minimum for a programme, weekly if you are actively campaigning or in a volatile category. Less than monthly and the trend line takes too long to mean anything. Within each run, repeat every prompt at least twice: in our 5 October 2026 sample, two runs of the same question minutes apart cited the same domains only 55% of the time.
Can this be done manually or does it need tooling?
Fifty prompts across four engines is a disciplined afternoon by hand, monthly, and the log structure on this page is enough. Tooling buys cadence, scale and competitor depth. Method first, tooling second: a tool that produces one blended score fails this method whatever it costs. Our comparison of AI visibility tools lists what each one tracks and charges.
Why do results differ between engines?
Each engine mixes live search, licensed feeds and its own weighting differently, so the retrieved set differs and the answer with it. In our 204-answer sample, 652 of the 762 domains cited were cited by one engine only, and 20 by all three. That variance is signal: it tells you which engine's trusted sources you are missing from.
What is a good non-branded visibility rate?
There is no universal benchmark, which is why the baseline and the competitor set matter. The useful question is your rate against your named competitors on your buyers' prompts, and which way it is trending. Categories vary too much for a global number to mean anything.
How do you know how your brand appears in ChatGPT answers in Great Britain?
Run the prompt set from a UK location, with ChatGPT's search on and no account history, and record the country in the log. Engines retrieve different sources for the same question asked from the UK and the US, so a US run does not describe what a British buyer sees. If your buyers are in both markets, keep two panels.
Start With a Documented Baseline
Find out what AI engines say about your brand, who they recommend instead, and what we'd change first.
Rather talk it through? Get in touch.