An AI mention is not a customer.

A screenshot proves that a brand appeared in one response under particular conditions. Keep the prompt, platform, date and cited URLs, then repeat a consistent set of questions. That gives the observation a context a single “visibility score” can hide.

A completed visibility log, not a vanity score

Here is a worked example log for one question: “Which inventory tools support multi-location stock control?” The numbers demonstrate a recording method, not measured brand performance.

Three mentions out of four runs is a 75% mention rate for this tiny test. It is not 75% market visibility, a stable ranking or proof of acquisition. One citation is a separate observation. Accuracy is another. Keep those measures separate.

Your log needs the exact prompt, platform, date, relevant account or location conditions, answer, named competitors and cited URLs. Repeat a fixed core of buyer questions; keep experiments in a separate group. Changing the question set changes what the score means.

Correct repeated factual errors before chasing more mentions. Keep referrals and enquiries beside the observations without claiming complete attribution.

Measure the observation, not an imaginary audience

An answer can mention you, cite somebody else and describe your product incorrectly, all at once. A single visibility score hides that awkward combination. The reporting has to preserve it. Track what appeared, where the information came from and whether the claim was right. Commercial evidence belongs beside those observations, not inside an invented conversion from citations to customers. The point is to identify a useful intervention, including when that intervention is correcting the brand rather than increasing exposure.

Build a question set around buyer decisions

Group questions by purpose: understanding the problem, comparing approaches, selecting a supplier and planning implementation. Include relevant markets and languages. Keep the exact wording in a log so changes in the test are not mistaken for changes in visibility.

For each observation, save the date, platform, account state where relevant, answer and cited pages. Record whether your business is mentioned accurately. A recommendation with the wrong service or geography is not an uncomplicated win.

Keep the metrics in separate columns

Microsoft’s AI Performance announcement describes citation reporting in Bing Webmaster Tools for supported AI experiences. Its reported coverage is not a view of every answer engine.

Design a reporting rhythm you can sustain

I would begin with a limited set of meaningful for the business questions, repeat it consistently and review patterns at a page and topic level. One answer changing overnight is not a strategic signal. Look for repeated misrepresentation, persistent citation gaps and pages that earn useful referrals.

Pair those observations with your existing analytics. Preserve identifiable referral sources when available, check the landing pages they reach and follow any resulting enquiries into the CRM. Missing referrer information means some journeys will remain unknown. Say so in the report.

Put the data to work

If you are cited for definitions but absent from comparisons, inspect your decision-stage content. If you receive visits that never progress, examine offer fit and the next step. If the system repeatedly describes you incorrectly, check whether your own pages contain contradictory descriptions.

Do not rewrite everything after a single disappointing prompt. Log a hypothesis, change a specific part of the information and keep an unchanged question set. Other factors can still affect the answer; this is observation, not a controlled ranking experiment.

A citation that brings unsuitable visitors needs a commercial content review, not a bigger visibility target. If reports disagree about the resulting enquiries, use the attribution decision framework before assigning revenue credit.

Catch a broken denominator before reporting a lift

Make two copies of your observation log. In the first, keep the original core question set. In the second, include the newly added questions. Calculate mention and citation rates separately, using completed observations as the stated denominator and reporting failed observations alongside them.

If the combined score rises while the unchanged core is flat, the apparent lift may come from the sample change. Report that explicitly. Do not erase unsuccessful runs or silently exclude awkward answers.

For an accuracy check, preserve the claim, the relevant current product information and the discrepancy. “Brand mentioned” and “brand described correctly” need separate columns. Otherwise a misleading answer improves your headline score precisely when it creates a new commercial problem.

What changed in first-party reporting in 2026

Bing’s 16 June 2026 announcement added Intents, Topics, Citation Share and Compare in preview. These extend citation analysis; they are not a universal AI ranking or revenue measure.

Use the available intent and topic views to ask whether visibility matches the questions suitable buyers actually ask. Keep the date range and coverage visible. A change in citation share deserves investigation, not an automatic budget increase. Compare it with the fixed-question log and observed enquiries; they answer different questions.

Learn from a published measurement adjustment

Semrush’s account of its own AI visibility work explicitly distinguishes an initial prompt set from a later, larger set and resets the baseline because the scopes are not directly comparable. That is the reporting discipline worth borrowing. The company’s percentages should not become targets for a different brand with different questions.

For your own programme, keep one stable core set and a separate exploratory set. The core supports a consistent view. The exploratory set lets you investigate emerging buyer language without quietly changing what the original measure means. Report both by intent group so informational questions do not obscure commercially important comparisons.

Build a spreadsheet that keeps the uncertainty visible

Use one row per observed answer. Include the exact question, platform, timestamp, language, whether the brand appeared, whether the description was accurate, any cited page and a note about material conditions. A separate summary tab can aggregate the observations, but the raw rows should remain understandable to someone who did not collect them.

Do not add an estimated reach figure unless the methodology and assumptions are defensible. Search volume for a related keyword is not automatically the number of people who asked an AI system your exact question.

Sample consistently without pretending you know every answer

A small team can start with a manageable schedule and a limited question set. Repeat the most important observations, especially when a change could influence a decision. Record variability rather than selecting the answer that best supports the story you want to tell.

If one platform produces a different response on each run, show a range or the proportion of your observations containing the brand. Explain the sample size. Three favourable answers out of five checks is an observed sample result; it is not proof that sixty percent of all users will see the brand.

Turn four kinds of findings into four kinds of action

A persistent factual error suggests a review of your own descriptions and genuine external information. A relevant citation to a weak landing page suggests a conversion review. Strong visibility for introductory questions but little for comparisons suggests a decision-content gap. Good referrals with poor qualification suggest a mismatch between the answer, page and offer.

These actions are more useful than pursuing a larger global score. They identify the work a team can do and the outcome it can inspect. The next report should show whether that work was delivered and what subsequent observations suggest, without claiming sole causation.

Make the observation repeatable before calling it a trend

Keep an observation record that another reviewer can understand without asking what you meant. Save the exact question, the date, the platform, language, available environment details and the completed answer. Record a brand mention, a cited URL and an accurate description as separate observations.

A mention without a link is not a referral. A citation with a wrong claim is not an uncomplicated win. An incomplete response is not automatically an absence. Those distinctions determine whether the next action is to correct a page, investigate a source or repeat a failed observation.

Keep the core question set unchanged when comparing periods. Add new buyer questions in a separate exploratory set. Otherwise an easier question mix can look like improved visibility. Report the numerator and denominator, not just a percentage: five citations across twenty completed observations tells the reader more than “25% visibility.”

This is a reproducible observation method, not a claim to measure every answer a platform produces. Link any actual referral activity to a separate website and CRM review. The useful decision is which inaccurate or missing buyer answer deserves attention, not whether a dashboard number can be made to rise.

The score can improve when the method gets worse

Removing unanswered prompts from a report can raise the percentage without changing a single successful answer. Adding easier brand questions can do the same. Preserve response failures and version the question set before calculating a trend.

When a repeated core question produces inconsistent answers, show the observations rather than selecting the most flattering result. Keep the platform and checking conditions attached to the rows. Variation is part of the finding.

A useful month ends with a decision supported by that finding: correct a recurring error, improve the cited destination or investigate a missing commercial answer. If the only action is “increase visibility,” the reporting has not yet explained what the team can change.

Keep the buying question in the observation record

Add a column for the decision the prompt represents: discovery, comparison, implementation or supplier selection. Keep the classification fixed during a comparison period. A change in visibility across a large set of broad definitions may matter less commercially than a persistent incorrect description on a few relevant buying questions. Keep both findings, but prioritize them differently. The record should show why the prompt belongs in the study, the market and language, the observed answer and the next investigation. It cannot reveal how many real buyers used that answer.

Sources and further reading

Microsoft Bing: AI Performance public preview, February 2026: First-party citation reporting and its coverage.

Microsoft Bing: new AI visibility insights, June 2026: Preview additions to citation analysis.

Google Search Central: AI features and your website: Google AI eligibility and reporting boundaries.

References checked 9 October 2026. Sources support the documented facts and tool behaviour; the recommendations and worked examples explain how I would apply them.