How to Check Your Brand's AI Visibility

How can I check my AI visibility?

Check AI visibility by testing a defined set of customer questions, saving the answers, and recording whether your brand is mentioned, recommended or linked as a source. Repeat comparable checks over time and separate results by engine. Use available publisher reports and referral analytics as additional evidence, not as interchangeable versions of the same visibility score. Keep paid placements apart from organic mentions; ads in AI answers explains where they now appear.

You can start manually. What matters is that another person could repeat your check and understand what the result means.

Asking "Do you know my company?" is not enough. It tests recognition after you supplied the name. It does not tell you whether a buyer would discover that company while researching a problem.

A quick AI visibility check you can do today

For a small first pass, select ten relevant questions and two AI surfaces. Run each question once, then repeat the same set on another day. That produces 40 observations if every run succeeds.

This is a manageable diagnostic sample, not a statistically representative estimate of all AI users.

Step 1: define the brand you are checking

Write down the official company name, product names, genuine aliases, owned domains and the competitors relevant to this business area.

Ambiguous names need human review. An answer mentioning "Square" as a shape should not count as visibility for the payments company. A reseller's page should not count as an owned-domain citation unless you deliberately include it in your ownership definition.

Step 2: choose questions that resemble buyer decisions

Use a mix of intents rather than ten variations of "best tools."

Keep branded and unbranded results separate. Branded questions can be valuable for checking misinformation, but they should not inflate a discovery score.

Step 3: record the conditions

For every observation, save:

• Exact question and prompt-set version.
• Date, time and time zone.
• Product surface, model or mode when visible, and whether search was used.
• Language and relevant location.
• Session context, including whether it was a new conversation.
• Available personalization settings.
• Full answer, source links and run status.

OpenAI documents that memory and location can influence search-query rewriting.[1] A fresh chat helps reduce conversational carryover, but it does not prove the absence of all personalization.

Likewise, an API response is an observation of that API configuration. Do not label it as a screenshot-equivalent result from a consumer app you did not test.

Step 4: classify what actually happened

Use separate yes/no fields:

1. Mention: Did the answer name the brand?
2. Owned citation: Did it link to a verified owned domain?
3. Recommendation: Did it suggest the brand as an option for the requested use case?
4. Accuracy issue: Did it make a material incorrect claim about the brand?

A source list containing your website is not automatically a recommendation. An answer can recommend your product while citing a third-party review instead of your domain.

Save ambiguous cases for review instead of forcing every answer into a positive or negative label.

Step 5: calculate rates with explicit denominators

For the manual panel above:

• Mention rate = valid answers mentioning the brand ÷ valid answers checked.
• Owned-citation rate = valid answers linking an owned domain ÷ valid answers checked.
• Recommendation rate = valid answers recommending the brand ÷ valid answers checked.

Count a brand at most once per answer for these rates. Track raw citation occurrences separately if you need them.

Also report scheduled checks, successful checks, missing answers and technical failures. If one engine had more failed runs, the remaining sample may be systematically different. A rate based only on successful answers should disclose that limitation.

One dataset can produce several different scores

Here is a constructed example, not CiteLadder customer data:

• 12 checks were scheduled.
• 2 failed technically.
• Of the 10 valid answers, 4 mentioned the brand.
• 2 linked its owned domain.
• Those 2 answers contained 3 owned citation occurrences among 15 total citation occurrences.

The resulting measurements are:

The two citation percentages happen to be equal here, but they describe different things. One asks how many answers linked the site; the other asks how much of the citation inventory belonged to it.

Using all scheduled checks as the mention denominator would produce 33.3%. That is not the same metric as 40% of valid answers. Report the convention instead of presenting either figure as a universal score.

The observation schema and worked example contain the inputs used in these calculations.

Worked example: share of tracked brand appearances

Consider an illustrative panel with 100 completed answers. Brand A appears in 30 answers, Brand B in 50 and Brand C in 20. A response may contain more than one brand.

Brand A's answer-level mention rate is 30 ÷ 100, or 30%. Under a rule that counts each tracked brand once per answer, its share of tracked appearances is 30 ÷ (30 + 50 + 20), also 30%.

If Brand B instead appears in 70 answers while the other counts stay the same, Brand A's mention rate remains 30%, but its share of tracked appearances becomes 30 ÷ 120, or 25%.

All numbers in this example are illustrative. The example shows why a report should state the competitor set, counting rule and denominator. Learn more about AI search share of voice.

Which free reports can help?

Google Search Console

Google's current dedicated generative AI report documents impressions for AI Overviews and AI Mode, with page, country, date and device breakdowns. Property-level and page-level totals may differ because they aggregate appearances differently.[2]

Use it for the Google surfaces it covers. It is not a report of every mention in ChatGPT or Gemini Apps. Do not relabel its documented impressions as visits or conversions.

Bing Webmaster Tools

Microsoft's AI Performance reporting covers supported Microsoft and partner experiences. Its June 2026 preview added intent, topic, citation-share and time-comparison views.[3]

Microsoft defines citation share against citations for the same grounding query and says it is not traffic share. Compare it with your own panel only after checking that the scope and denominator align.

Your analytics platform

Referral reports can show identifiable visits from AI services; our AI referral traffic statistics compare how much each assistant sends. They cannot show a person who read an answer and never clicked, and attribution can be lost or changed along the journey.

Use landing-page and conversion analysis to ask whether the traffic was useful. Do not multiply a mention rate by website sessions to manufacture an estimate of AI impressions.

What should an AI visibility report contain?

A useful one-page report answers five questions:

1. Scope: Which products, markets, questions and engines were checked?
2. Coverage: How many observations were valid, missing or failed?
3. Presence: Where was the brand mentioned, cited or recommended?
4. Change: What moved within the same prompt panel and conditions?
5. Action: Which content, accuracy or access problem deserves investigation?

Include a few actual answer examples and their sources. A percentage without inspectable evidence is hard to challenge or improve.

Separate competitor mentions from competitor-owned citations. A publisher can cite your competitor's research without recommending its product, just as it can recommend yours through an independent source.

How often should you check?

Choose a cadence that fits the decision. A weekly review can be practical for an early content programme; a time-sensitive campaign may need more frequent observations. These are workflow choices, not universal statistical requirements.

Retain a fixed core question set. Add new questions as a separate cohort, rather than merging them into an old trend without explanation.

For each content update, record a hypothesis such as: "This page will answer the integration constraint that was missing from the current sources." Then review whether the observed answer and source pattern changed. A change is evidence to investigate, not automatic proof that the edit caused it.

Common mistakes when checking AI visibility

Counting brand-led prompts as independent discovery

"What makes Brand X the best?" pushes the answer toward Brand X. It may test persuasion or accuracy, but it is a poor baseline for unprompted category visibility.

Treating every link as the same kind of citation

Gemini's double-check links, for example, need a different label from sources displayed with the original answer. Our citation tracking guide explains a consistent classification approach.

Averaging different panels together

If one engine has 100 generic prompts and another has ten niche prompts, an unqualified combined percentage can hide the difference. Publish per-engine and per-intent results first.

Treating no appearance as a diagnosis

Absence tells you what happened in the observation. It does not tell you whether the cause was access, retrieval, relevance, competition or answer variation. Investigate before prescribing a fix.

Manual checks or an AI visibility platform?

To build an inspectable baseline from collected answers, explore CiteLadder AI Visibility. For identifiable visits reported by GA4, use AI Referral Analytics; those sessions are separate from answer visibility.

Manual checks are useful for validating your definitions and inspecting answer quality. A platform becomes more useful when the cost of retaining and comparing evidence exceeds the cost of automation. For repeated prompt runs, see the LLM rank tracker guide.

Before choosing one, ask whether it preserves raw answers, identifies the engine and run conditions, exposes source URLs, separates mentions from citations, and lets you export the underlying observations. See our AI visibility platform comparison for a buying framework.

CiteLadder's workflow centres on retained answers, prompt/run context and inspectable source references.[4] Use those records to investigate a result, not just to report a score.

Explore AI citation tracking and then learn how to improve AI visibility based on what your baseline reveals.

Sources

1. OpenAI, Searching the web with ChatGPT.
2. Google, Generative AI performance report.
3. Microsoft Bing, AI visibility reporting updates, 16 June 2026.
4. CiteLadder, product information reviewed 6 October 2026.