How to evaluate AI-search visibility without pretending there is one rank.

Traditional rank tracking assumes a repeatable list of results. AI answers do not offer one universal list. The output can change with the engine, model, product mode, location, account state, source refresh, and the way a question is phrased. A single screenshot cannot establish a permanent position.

The defensible unit is an observed question-and-engine cell: one exact question, asked in a recorded context, on a recorded date. Preserve the answer state and source URL, then repeat the same cohort later. Describe the result as observed coverage, not “AI rank.”

Start with real buyer decisions, not a pile of near-duplicate keywords. Confirm the business's market, services, audience, pricing or quote rules, constraints, named alternatives, and official entity details. Then write six questions in each of four groups.

Replace every bracketed field with approved facts. Save the exact wording as cohort version 1. Do not rewrite weak questions between monthly runs just to improve the result. If the offer or market changes enough to require a new cohort, label it a new baseline.

Name the engines and product modes before the run. Use the same approved questions within a bounded observation window. Manual review matters because the evidence is the rendered answer and its links, not a simulated API response or a search snippet from another product.

Record these fields for every cell

• Exact question and cohort version
• Engine, product mode, and model when shown
• Date, time window, location, and login state
• Cited, mentioned, absent, or N/A/blocked state
• Owned URL cited, when present
• Answer excerpt or screenshot reference
• Reason for N/A or blocked status
• Reviewer and review notes

If a login wall, regional limitation, rate limit, safety response, bot challenge, or product outage prevents a comparable answer, mark the cell N/A or blocked. Never silently turn an unobserved sample into “absent.”

Keep counts and denominators visible. A practical report can state:

Report N/A and blocked cells beside those rates, never inside the denominator. Also list the actual cited pages and source links. A mention without a link is useful entity evidence, but it is not a citation and should not be relabeled as one.

Bing's AI Performance report can add aggregate citation and cited-page evidence for supported Microsoft experiences. Bing explicitly says those metrics do not indicate placement, page importance, authority, or rank within an answer.

Use the baseline to find questions the site should answer better. The goal is not to repeat a phrase until a model notices it. The goal is to publish a page a buyer would choose to read and an engine can quote without stripping away a critical condition.

For every substantial update, retain the old and new page state, the buyer question it addresses, the supporting source, the approval, and the publish date. Avoid doorway pages, hidden text, keyword spam, unsupported superlatives, and mass-produced pages that add no new help.

Visibility is an upstream observation. A citation is not a lead, and a lead is not revenue. Keep those claims separate until evidence connects them.

Use explicit source fields in the contact or CRM record, preserve landing-page and referrer data where consent and tooling allow it, and ask a plain “How did you hear about us?” question when useful. Self-reported “ChatGPT” or “Perplexity” is evidence of the buyer's recollection, not proof of the exact answer or page that caused the visit.

Month one creates the technical and question baseline. Months two and three repeat the cohort and add substantive, approved page changes. At the 90-day review, compare like with like and decide whether the observed signal and business evidence justify continuing, changing scope, or stopping.

Do not claim causation from a before-and-after count alone. AI products, source indexes, competitors, and answer generation all change. The review is a decision aid grounded in observed evidence, not a promise that citations, traffic, or leads will increase in 90 days.

Crawler names, product behavior, and reporting features can change. Check the live publisher documentation before changing access controls or making a measurement claim.