TL;DR
AI brand visibility is not one ranking. It is what a brand looks like in a defined sample of AI answers: whether it is named, recommended, described correctly, or linked as a source. We suggest testing relevant unbranded questions across platforms, recording the answers and citations, and comparing like-for-like samples over time. A single check is a starting observation, not proof of how often customers see your brand.
When someone asks an AI assistant to recommend a provider, your website can appear as a source, your company can appear in the answer, both can happen, or neither can. Those outcomes call for different responses. A brand named for the wrong service has a positioning problem; a useful page cited without a brand mention has a different kind of visibility.
In this guide, we'll define the signals, set up a manageable measurement baseline, and show how to turn what you observe into better pages and more credible evidence. The aim is not to manufacture an all-purpose AI score. It is to make the next marketing decision less speculative.
Table of contents
- What does AI brand visibility mean?
- Why the answer changes across platforms and prompts
- Build a prompt set around customer decisions
- Create a baseline you can actually repeat
- Turn the observations into a readable report
- Diagnose the gap before producing another page
- Prioritise improvements without promising a citation
- AI brand visibility measurement checklist
- FAQ
- Start with a defensible snapshot
- Sources
What does AI brand visibility mean?
AI brand visibility is your brand's observable presence and portrayal in AI-generated answers to questions relevant to your business. Its unit of analysis is an answer to a particular question, on a particular platform, under particular conditions. Traditional search rankings describe a page's position in a list; they do not, by themselves, tell you whether an assistant named your business or linked to it in a generated response.
We separate five signals because each answers a different question:
| Signal | What to record | What it does not prove |
|---|---|---|
| Mention | Is the correct brand named in the answer? | That the assistant endorses it. |
| Recommendation | Is the brand presented as a relevant option for this need? | That the customer clicked or bought. |
| Description | Are its category, capabilities, audience, and limitations accurate? | That every future answer will agree. |
| Citation | Does a source link point to the brand's domain or a relevant third-party page? | That the cited brand was named or recommended. |
| Competitive presence | Which other brands appear in the same sampled answers? | Market-wide share of voice. |
A citation can be a link to your page even when your company name never appears in the prose. Conversely, a recommendation might name you without linking to your site. Record the destination of each source link, not just the presence of a citation icon. If the assistant cites a comparison site that mentions you, that is not the same as a citation to your own domain.
Share of voice needs a stated denominator. For example, you could count the number of sampled answers mentioning your brand divided by all answers mentioning at least one brand in a defined competitor set. Another team might divide your brand appearances by all appearances of brands in that set, including multiple brands in one answer. Both can be useful, but they produce different percentages. Label the method and competitor set so the trend is interpretable.
Why the answer changes across platforms and prompts
There is no single AI results page that every customer sees. ChatGPT may search the web for current information and show source links, while Google says its AI Overviews and AI Mode can use different models and techniques and therefore display different answers and links. Google also notes that an AI Overview will not trigger for every query. Compare each surface separately rather than averaging them into an unexplained universal rank. OpenAI explains ChatGPT web search and its sources; Google documents how its AI search features work.
Conditions matter too. Change the question from “payroll software for a small UK business” to “alternatives to an enterprise payroll suite” and you have changed the buyer need. Add a city, switch search on or off, or run a check at a later date, and you have changed the observation. OpenAI says location information can affect local results and warns that search citations can be incomplete or incorrect, so open the links instead of accepting them at face value. OpenAI's search documentation is a useful reminder to log those conditions.
This variation is measurable, not merely theoretical. A 2026 research preprint on citation uncertainty repeated queries across three generative search platforms and three consumer-product topics and found substantial differences among cited domains across samples. That limited study does not establish the right sample size for your business, but it does make a practical point: treat a one-off result as an observation, not a stable property of the brand.
Build a prompt set around customer decisions
If you test only “What is [brand name]?”, you mainly learn how the system describes a brand it has already been told to discuss. That is a useful accuracy check, but a weak test of unprompted discovery. For visibility measurement, start with the questions a buyer could ask before knowing your name.
Imagine an illustrative UK payroll provider called Harbour Payroll. Its initial prompt set could include:
- Category: “Which payroll software is suitable for a small UK company?”
- Problem: “How can a small business reduce manual payroll corrections?”
- Comparison: “How do payroll platforms for a growing UK team differ?”
- Alternative: “What are alternatives to our current payroll provider for a team of 40?”
- Recommendation: “Which payroll services should a UK hospitality business shortlist?”
These are examples, not measured search volumes or real recommendations. Replace vague wording with the language customers actually use in sales calls, support requests, site search, and relevant search queries. Keep prompts that are genuinely within your offering, and mark informational questions separately from requests to name vendors. An absence from a how-to answer is not automatically a missed recommendation.
If you operate in more than one region, keep a separate prompt group for each market. Add brand-name and misspelling checks as a diagnostic group, not as evidence that unknown buyers are discovering you. The GrowthScout guide to keywords and prompts explains why a search keyword and an AI question can reflect the same customer need without having interchangeable demand estimates.
Create a baseline you can actually repeat
Choose a fixed set of relevant prompts, the platforms you want to observe, and a cadence your team can sustain. One illustrative workload is 20 prompts on two platforms, with each prompt run twice per platform in separate sessions: 80 answer observations. Twenty is a planning example, not a statistically validated minimum. If your category is broad, divide the prompts into intent groups instead of adding many near-duplicates of one question.
For each observation, save the exact prompt, platform and interface, date, market or location setting, whether web search was available or used, answer text or capture, visible sources, and any errors. Note the brand names and domains you are matching. A failed response is an error, not a no-mention. With a common name, verify that the answer is about the right company before counting it. A worksheet can also flag whether the cited page actually supports the statement attributed to it.
Run the same core prompt set next time under comparable conditions, but keep new exploratory questions in a separate group. You can change your main set when buyer behaviour changes; just record the change so an increase in mentions is not mistaken for an improvement caused by easier prompts. Keep the raw answers alongside your summary: when a percentage changes, the examples explain what changed.
A first pass should also distinguish a snapshot from a trend. For example, GrowthScout's public AI visibility checker tests three category-based questions once on each of the web versions of ChatGPT and Gemini, and reports matches and sources under stated conditions. That is a useful initial look, not a comprehensive estimate of all buyer questions or a substitute for repeated, comparable observations.
Turn the observations into a readable report
Start with denominators and examples, then calculate rates. Suppose all 80 observations in the illustrative baseline returned usable answers. If your brand is correctly identified in 24, mention rate is 24 ÷ 80 = 30% for that sample. If nine include a link to your domain, domain citation rate is 9 ÷ 80 = 11.25%. Do not assume those nine answers also mentioned the brand; check the overlap. If some runs fail, use the number of usable answers as the denominator and report the failures separately.
We recommend a compact report with these fields:
| Measure | Definition for this report | Diagnostic use |
|---|---|---|
| Mention rate | Usable answers that name the correct brand ÷ usable answers | Find missing categories or prompt types. |
| Recommendation rate | Usable answers presenting the brand as a relevant option ÷ usable answers eligible for a recommendation | Separate neutral references from endorsements. |
| Domain citation rate | Usable answers linking to your domain ÷ usable answers that display source links | See whether your pages appear as sources where links exist. |
| Description accuracy | Correct descriptions ÷ answers that describe your brand | Identify wrong claims or dated positioning. |
| Competitor presence | Answers mentioning each named competitor within the same prompt set | See who appears when you do not. |
You may also calculate a share-of-voice measure, but publish the competitor set and counting method beside it. Segment rates by platform, market, and intent before combining anything. An overall number can hide a strong showing on branded questions and an absence from unbranded category recommendations.
Review source types as well: your site, independent reviews, directories, news, community discussions, and competitors' own pages play different roles in a buyer's evaluation. Open a sample of cited pages and note whether they support the relevant answer, how recently their facts were checked, and whether your business is represented accurately. Keep an eye on visits or enquiries from identifiable AI referrals where your analytics can show them, but do not equate a citation with a click. Google says visits from its AI Overviews and AI Mode are included in Search Console's overall Web performance data, so that report is not itself a complete count of your brand mentions in AI answers. Google's AI features documentation explains the reporting context.
Diagnose the gap before producing another page
A low mention rate is a symptom, not a diagnosis. Read several missed answers next to their source links and compare the source's claims with what your own pages actually say.
Relevant competitors appear, but you do not. Check whether your category pages clearly state whom you serve, what problem you solve, and which constraints matter. If the answer repeatedly draws on a comparison or specialist directory, review whether your public profile there is absent, incomplete, or out of date. An omission in a tiny sample is not evidence that you need to appear on every directory.
The brand appears, but the description is wrong. Find the inaccurate claim and the pages repeating it. Update your own product and about pages with clear, verifiable wording; where possible, request factual corrections to relevant third-party listings. Record the answer and source so you can see whether the description changes later. Avoid claiming you can directly edit a model's knowledge by changing one page.
Your site is cited, but the brand is not named. The page may be supplying information without presenting a clear connection to the provider behind it. Check whether the page identifies its author or business where helpful, answers the specific question, and makes its evidence easy for readers to verify. Do not add a brand name to every sentence or strip out useful independent information merely to chase mentions.
You are named, but the supporting link goes elsewhere. Open the cited page. It may be an appropriate independent review, or it may contain outdated details that need correction. A citation to a third party is not inherently a failure; it can show which outside material readers and systems may encounter when evaluating you.
If none of your useful pages is accessible, check technical fundamentals before rewriting prose. Google says eligible supporting links in its AI features must come from indexed pages that can appear in Search with a snippet; it also recommends crawl access, findable internal links, and important information in text. That guidance applies to Google Search, not automatically to every AI product. Google's guidance for AI features and websites sets out those requirements.
Prioritise improvements without promising a citation
We use three questions to select the next piece of work: How important is this buyer question? How strong is the evidence of a gap? How feasible is the fix? A recurring incorrect product description across relevant prompts deserves attention sooner than a single omission on a marginal question. A broken or outdated primary page is often easier to fix than a missing independent source.
Choose a specific action tied to the diagnosis. Clarify an existing service page if the offer is poorly defined. Publish a useful comparison or worked example if customers need criteria that your site does not explain. Add documented proof, methodology, or first-hand evidence where a claim currently has none. Correct a relevant third-party listing if it misstates the business. Then rerun the same prompt group and look at both the rates and the actual wording.
Improvement after an edit does not establish that the edit caused it. Other sites, platform behaviour, model updates, and sample variation can change answers. Even Google's eligibility guidance does not guarantee appearance. Use repeated observations to decide whether to keep investigating, not a single favourable response to declare victory. Google's AI features guidance and the citation-variability study both support that caution for different reasons.
AI brand visibility measurement checklist
Before sharing a visibility figure, we check:
- Is the brand and its domain unambiguous, including relevant aliases?
- Do the prompts represent category, problem, comparison, alternative, and recommendation needs without planting the brand name?
- Have we recorded platform, market, date, search conditions, failures, and the raw answers?
- Are mentions, recommendations, descriptions, and domain citations counted separately?
- Can we explain each denominator and the brands included in any share-of-voice calculation?
- Have we opened the cited pages and checked important claims rather than trusting citation labels?
- Does the proposed content or correction address a specific repeated gap, and will we retest comparable prompts?
FAQ
Does zero mentions mean AI never shows my business?
No. It means none of the usable answers in your tested set mentioned the correct business. Report which prompts, platforms, and dates you tested before drawing any wider conclusion. If your brand name is shared by other businesses, manually review ambiguous matches.
How often should we measure AI brand visibility?
Choose a cadence that lets you compare like with like, such as a monthly review of the same core prompts, and make an additional clearly labelled check after a substantial positioning or site change. There is no universal interval that removes answer variability. For a small set, read the individual answers as carefully as the percentages.
Can we use SEO rankings or keyword volume instead?
Use them as complementary signals, not substitutes. Search demand can help select relevant topics, and technical SEO helps eligible pages get discovered on Google. Neither metric tells you whether a sampled AI response recommended your brand or cited your domain. Keep the keyword and prompt measures in separate columns when you report results.
Start with a defensible snapshot
If you have no baseline yet, try our free AI visibility checker to see the specific ChatGPT and Gemini answers it collects and the source links it returns. Then broaden the prompt set to match your actual buyers, save the underlying observations, and choose one improvement you can test. We would rather help you understand five meaningful missed answers than celebrate an unexplained score.