Introduction
To audit your visibility in AI assistants, ask the same realistic customer questions across the assistants and markets that matter to your business, then record whether your brand is named, described accurately, and cited. A single answer is only a snapshot: assistants can return different sources and wording from one run to the next, so useful measurement depends on a consistent question set and repeated checks.
This audit is about what happens when an assistant searches or grounds an answer in current web sources. It does not prove what every language model knows from training, and it cannot guarantee what an individual customer will see. Treat it as a way to find discoverability problems and track evidence over time.
First, define what you mean by AI visibility
AI visibility is not one metric. A business can be mentioned without a link, cited without being recommended, or receive visits without appearing in every answer. Separate these outcomes before you start:
- Mention: the answer names your company.
- Description: the answer correctly explains what you sell, who you serve, and where you operate.
- Recommendation: the answer presents your company as a possible fit for the question.
- Citation: the answer links to your website or another source about your company.
- Referral: a person visits your site from an AI assistant.
- Business outcome: that visit produces a meaningful action, such as a lead or purchase.
These measures answer different questions. A citation is evidence that a page was used as a source; it does not necessarily mean the brand was recommended. A mention can happen without a website link. Referral sessions show that some people clicked through, but they do not show how often your brand appeared in answers that nobody clicked.
Step 1: Choose the assistants, market, and business scope
Start with the surfaces your customers actually use. Depending on your market, this could include ChatGPT search, Google AI search experiences, Gemini, or another assistant with web search or grounding. Record the product and mode you tested. A web-enabled answer is different from a response generated without live search.
Set the country and language as well. Recommendations can vary by market, language, and local availability. If you serve several regions, test each as a separate scope instead of mixing results into one score.
Write down the business facts you want answers to get right: company name, product or service, ideal customer, important use cases, service area, and meaningful differentiators. This is your reference for checking accuracy later. Keep it factual and concise; the audit is not a test of whether an assistant repeats your preferred marketing copy word for word.
Step 2: Build a fixed set of customer questions
Use questions a potential customer might ask before choosing a solution. Do not test only your company name. Branded searches can show whether an assistant knows your business, but they say little about discovery by someone who has not heard of you.
Include a balanced set of intents:
- Recommendation: “What are good options for [need] for a [type of business]?”
- Problem: “How can a [type of business] solve [specific problem]?”
- Comparison: “What should I compare when choosing [category of provider or product]?”
- Local or audience-specific: “Who serves [audience] in [location]?”
- Branded: “What does [company] do, and is it a fit for [use case]?”
Make questions specific enough to represent real needs, but do not write them to force your company into the answer. A question such as “Why is Reelistic the best agency?” is not a neutral discovery test. Keep the exact question wording, language, country, and assistant mode in a saved log. You can add new questions later, but preserve a stable core set so results from different dates can be compared fairly.
Step 3: Check whether the relevant pages can be found
Before interpreting an absent mention as a content problem, check the basics. The pages that explain your offer should be publicly accessible, return a successful response, and not accidentally block indexing with a noindex directive. Review your robots rules, hosting or CDN restrictions, and any access controls that might prevent crawlers from reading those pages.
The required crawler depends on the surface. OpenAI documents OAI-SearchBot for surfacing websites in ChatGPT search and lists GPTBot separately for model training. A site can control those crawlers separately. Google says its AI search features depend on the same general Search eligibility requirements: a page needs to be indexed and eligible to appear with a snippet. Allowing a crawler makes content accessible; it does not promise that a page will be indexed, cited, or shown in a particular answer.
Start with the pages a person would need to verify your business: homepage, product or service pages, industry pages, pricing or process pages where appropriate, and contact or location pages. Check that each page states its subject clearly and that the important facts are present in visible page content. See the Google Search guidance for AI features, Google's technical requirements, and OpenAI's crawler documentation for current platform-specific details.
Step 4: Run the questions and keep a record
For each question, record the date, country, language, assistant, mode, exact prompt, whether an answer was returned, and the answer's cited links. Save enough of the response to evaluate the results later. If the assistant or search feature is unavailable, mark that row as unavailable rather than as a zero.
Use the same process on each run. Do not change the prompt to make an answer more favorable, and do not combine results from different countries or assistant modes without labeling them. If a search returns no useful result, repeat it once if the product appears to have failed to load; otherwise record the empty or unavailable response as such. That distinction keeps technical failures from looking like poor brand visibility.
Answers can vary from run to run. Repeat the fixed set on a regular schedule and look for patterns across questions and dates. Avoid treating one response as a stable ranking or a promise of what all users will see.
Step 5: Score what each answer actually does
Use simple fields rather than one opaque “AI visibility” score. For every answer, note:
| Field | What to record |
|---|---|
| Brand named | Yes, no, or unclear |
| Recommendation | Recommended, listed as an option, or not included |
| Description accuracy | Accurate, partly accurate, incorrect, or not enough information |
| Website cited | Whether a link to your own site appears |
| Supporting page | Which of your pages was linked, if any |
| Other sources | Which external sources were cited |
| Competitors | Which alternatives were named |
| Answer status | Returned, empty, or unavailable |
Then calculate a few transparent summaries. For example, mention rate is the number of returned answers that name your brand divided by the number of returned answers in the same test set. Citation rate is the number of returned answers that link to your website divided by that same denominator. State the denominator and time period whenever you report a rate. Exclude unavailable assistant runs; do not silently count them as misses.
You can also compare results by question intent, assistant, or market. This shows whether you are visible for a particular use case or in a particular region, instead of hiding the variation in one overall percentage. With a small question set, show the counts alongside percentages—for example, “named in 3 of 10 returned answers”—so a reader can see how much evidence the number represents.
Step 6: Check accuracy and source quality
When an assistant gets something wrong, compare the answer with the current facts on your site and with the sources it cited. It may have used an old page, a directory listing, an outdated review, or a third-party article. Record the incorrect claim and its apparent source before deciding what to change.
Make your own key pages explicit and current. State what the product does, who it is for, the problems it addresses, and any important limits or service areas. Use consistent names for the company and its offers. If a fact has changed, update the page that owns that fact and check whether the same information is stale elsewhere.
Also review the sources that appear repeatedly. Accurate business profiles, relevant independent coverage, and useful customer information can help people and systems verify what you do. Do not create fake reviews or low-quality pages simply to manufacture mentions. Google recommends content that is original and useful to people; its AI guidance does not require special AI-only markup or a separate file to make a site eligible for AI features.
Step 7: Measure visits separately
Use your analytics platform to review visits that arrive from AI assistant referrals, the pages those visits land on, and any key actions they complete. Compare equivalent date ranges and account for referral traffic that analytics may not identify consistently. If you connect Google Analytics 4, group the relevant assistant sources and compare sessions and key events with the prior period.
Keep referral measurement separate from answer testing. Many people do not click a citation, and some assistant visits may not retain a clear source label. A rise in referral sessions is useful evidence of visits, but it is not a complete measure of how often the business is mentioned or recommended. Likewise, a citation appearing in an answer does not prove that it generated a visit or sale.
How Reelistic approaches the audit
Reelistic's SEO and GEO audit uses a saved set of brand-specific and brand-free questions, so recurring checks can use the same wording. It tests supported AI answer engines and Google search results for brand mentions, position in the answer, citations, competitors, and whether the description is accurate. It can also review AI referral sessions and key events when Google Analytics is connected. Results that depend on a changing answer are treated as observations from that run; comparisons are most useful when the question set stays the same.
This is one part of a broader growth system. The aim is to connect discovery signals to the pages and campaigns that can improve them, then review whether those changes produce better qualified visits and business outcomes. For context, see how to build a cross-channel growth system and what to measure across SEO and GEO.
A simple reporting format
A useful report does not need a complicated dashboard. Capture the scope, the date, the stable question set, returned versus unavailable answers, mention and citation counts, accuracy issues, recurring sources, and referral activity if analytics is available. Include a few representative answer examples with their citations, and list the pages or facts you plan to improve.
On the next audit, compare the same questions and markets first. Then note any new questions separately. If results shift, check what changed: website access, page content, external sources, the assistant's retrieval, or the question set. The audit is a feedback loop for investigation, not proof that one edit caused an answer to change.
Related Reading
Frequently Asked Questions
How often should I audit AI assistant visibility?
Run the same saved questions on a regular schedule, such as monthly or quarterly, and repeat sooner after a meaningful change to your website or business facts. The right interval depends on how quickly your market and content change. Always compare equivalent question sets and markets.
Does being mentioned mean an AI assistant cited my website?
No. An assistant can mention your business without linking to your website, or cite your website without recommending your business. Track mentions, citations, referrals, and outcomes separately.
If my brand does not appear, does that mean AI cannot find my website?
Not necessarily. The assistant may have used a different source, interpreted the question differently, or not returned the same answer on that run. Check site accessibility and indexing, inspect the sources shown, and repeat the fixed question set before drawing a conclusion.
Should I block GPTBot if I want to appear in ChatGPT search?
OpenAI documents OAI-SearchBot for ChatGPT search and GPTBot separately for model training. Review the current crawler documentation and your own policy before changing access. Allowing OAI-SearchBot can make pages available for search, but it does not guarantee a mention or citation.
What is a good AI visibility score?
There is no universal score that predicts leads or sales. Report counts and rates with the question set, assistant, market, date range, and denominator. Use them to find patterns and decide what to investigate, then evaluate visits and business outcomes separately.