AI Visibility Scores: How to Check AI Citations Yourself

Several different events often get grouped under the phrase AI visibility. An answer might name your company without a link. It might link to a page that never mentions your company. It might describe your services incorrectly. A visitor might arrive from an AI product after reading several answers. These events carry different evidence and should have different labels in your report.

Keep these labels separate even when the same event produces several of them. A citation that gets your service area wrong needs a correction task before anyone puts it in a chart. A useful referral to a narrow service page may deserve attention even if your brand mention count stayed flat. Clear definitions keep your team from chasing the easiest number to collect. We label our own numbers the same way. Our records for Garrison, the microsite software from Trojan Digital Marketing, show four metal roofing inquiries from 203 Indiana sites from June 29 through August 31, 2026. We call them inquiries, since they don't show which became jobs.

For each manual check, record the exact question, product used, visible model or mode when available, date, location assumptions, and whether the session carried earlier conversation context. Save the answer and its links. Record whether you were signed in if that could affect the experience, without storing unnecessary account information. Mark anything you cannot establish as unknown.

Then verify the cited destination. Does the link open? Does it lead to the page named in the report? Does the page actually support the statement? A citation that points to an outdated service page is different from an accurate citation to a current explanation. A brand name in a source title is also different from a recommendation within the answer.

Your log should include a short human judgment alongside the raw observation. Suitable judgments include accurate, partly accurate, unsupported, outdated, or not enough context. Add the specific sentence that needs investigation. This turns a collection of screenshots into a working improvement queue, while preserving the distinction between what the system said and what your reviewer concluded.

Use a deliberately small question set that reflects your customers' decisions. Include an unbranded category question, a service and location question where geography matters, a comparison question, and a factual question about your business. Keep discovery questions separate from branded questions. Asking an assistant to describe a named company tests a different task from asking it to recommend providers.

For a hypothetical commercial cleaning company, a useful set might cover cleaning an occupied office, coordinating after hours access, comparing recurring service with a one time clean, and confirming which locations the company serves. The set should come from genuine sales conversations. It should not expand into hundreds of near identical prompts solely to make the report appear comprehensive.

Freeze the core set for a review period. Add experimental questions in a separate section. If you change every question between reports, you lose the ability to explain whether the answers changed or the test changed. Log repeated checks as repeated samples. They don't show every answer every customer gets.

A score is a model of selected evidence. Before treating it as a target, ask which products it covers, how questions were chosen, how often checks run, and how mentions and linked citations are weighted. Ask whether your competitors were selected manually, whether duplicate answers are counted, and whether the calculation changed since the previous report.

Request a sample behind the number. A vendor should be able to show enough underlying observations for you to understand the result. You do not need every proprietary implementation detail. You do need to know whether the score represents observed answers, a technical checklist, a prediction, or an opaque combination. Each can help, but they answer different questions, so we ask for the sample first.

Do not compare scores across tools as if they share a unit. A change from one vendor's scale to another can create an apparent improvement without any change in customer experience. Preserve the tool name, methodology version, and sample dates beside the result. When a methodology changes, begin a new comparison series or clearly mark the break.

Google's dedicated generative AI performance report provides impression data with page, country, device, and date views. Google announced global rollout by August 31, 2026. This provides platform evidence of visibility; it does not provide a complete prompt level citation history or a record of qualified leads. See the official report documentation and rollout announcement.

Label each export with its report type. A standard Web search export alone does not isolate generative AI performance. Likewise, do not add a dedicated AI impression total to an overall total that already includes that activity. When a report is absent or a field is unavailable, leave a gap in the record and state the limitation.

Website analytics and sales records answer the next set of questions. Our explanation of separating AI referrals from search traffic helps prevent platform impressions from being mixed with site sessions. That distinction matters when a business needs to allocate a limited marketing budget.

Prioritize corrections that matter to a buyer: wrong availability, wrong service area, missing prerequisites, unclear pricing conditions, or an unsupported credential. Identify the public page you control that should explain the fact. Update it after the business owner confirms the information. Record the publication date and recheck the same question later, without promising that a particular system will adopt the change.

For broader content improvements, use our search and AI visibility services as a framework for connecting discoverability, useful pages, and measurement. If the facts exist but are buried in awkward navigation or unreadable layouts, the problem may belong in a website improvement project. The evidence should determine the work, rather than a tool's generic recommendation list.

A monthly review can be brief. Check whether previously identified inaccuracies remain, whether new relevant pages appear in observed citations, and whether recognizable referrals lead to meaningful on site actions. Document competing explanations: a service launch, a seasonal change, a tracking repair, or a different prompt sample. An honest report can show progress while leaving causation unresolved.

Use the AI search content measurement plan to assign owners and review dates. End each review with a decision: correct a fact, improve a page, investigate a measurement gap, or leave a stable page alone. A useful score earns its place by helping you make that decision. The saved evidence is what lets another person check whether the decision was sound.