AI visibility measurement should show where a B2B company appears in relevant answers, which visits reach its website and what happens to those enquiries. Each measure needs its own definition. A citation can establish presence in an answer. A qualified enquiry requires evidence about a real prospective customer.
For a marketing leader, the useful monthly question is whether the evidence calls for a page change, a further check or more time. A single visibility percentage rarely provides enough detail to make that decision.
This guide sets out a repeatable measurement method, with a worksheet that can sit beside the B2B website performance scorecard. The numerical examples are illustrative. They are not client results or industry benchmarks.
What should an AI visibility report measure?
A useful report separates observed answers, website visits and commercial outcomes. Keep these in distinct sections so the same number cannot quietly change meaning.
Answer presence records whether the company is mentioned, linked or recommended in a defined set of answers. Record a link to the company’s own website separately from a third-party article mentioning it. A competitor comparison that criticises a product still counts as a mention, so inspect the surrounding answer.
Referral visits records sessions with an identifiable AI source. Break them down by source and landing page. These show activity that reached the website; the number excludes people who read an answer and take no measurable website action.
Qualified enquiries records submitted requests that meet an agreed definition, such as a relevant company with a need the business can serve. Record rejected, duplicate and unresolved enquiries separately. The team receiving the leads must review their fit.
The resulting report might show wider answer coverage, unchanged referral visits and one enquiry awaiting review. That is a usable description of progress and uncertainty. Turning it into an estimated revenue gain would require more evidence.
How do you build a repeatable prompt panel?
A prompt panel is a fixed set of questions checked across named AI products. Start with questions from sales conversations, search queries and customer research. Save where each question came from and the buying decision it represents.
Separate questions about a problem, a supplier comparison and the company itself. A branded question already contains the answer you hope to observe. Keep its results apart from questions where the assistant must discover the company.
For each question, record:
- Exact wording and a stable question ID.
- Source of the question and intended buyer.
- Country and language being evaluated.
- AI product and mode, including whether web search is used.
- Account context, collection method and date.
- Planned repetitions and the saved answer for each run.
A prompt mentioning Ireland does not prove the product served an Irish user experience. Record actual location settings where available and disclose where location could not be controlled. API responses and consumer interfaces also need separate labels when their collection methods differ.
Keep the questions stable between reviews. Add new questions to a separate list. When the main panel changes, give it a new version and keep the old set so results can still be compared.
Repeat checks help reveal whether a result recurs. Choose how often to run each check before collecting results and apply that schedule consistently. Three repeats, for example, are an operating choice; they do not make the sample representative of every buyer using the product.
Define citation coverage before calculating it
For this worksheet, citation coverage means completed answer runs containing at least one link to the company’s domain, divided by completed eligible answer runs.
Count each answer once, even if it links to three pages on the same domain. Store those individual URLs for analysis. This avoids giving a longer answer more weight simply because it has more links.
An illustrative UK panel contains 12 questions checked three times in one product, giving 36 planned runs. Four fail technically. Ten of the 32 completed answers cite the company.
Report 10/32 completed answers cited the domain, or 31.3%, with 4/36 checks unavailable. The four failures remain visible as missing evidence. They are neither successful citations nor confirmed absences.
A completed answer without a citation is an observed absence under this definition. A failed check is missing. If the product chooses not to search, record that too. Decide beforehand whether the main measure includes those answers, then keep the rule consistent.
Show coverage per product and market before any total. An equal number of questions per product produces an equal-weight panel, not an estimate of how customers divide their usage between products. Do not label the combined percentage market share.
Which platform reports can support the panel?
Platform reports provide observations from their own systems. Keep their metric names and scope alongside the panel measure.
Google introduced dedicated Search Generative AI performance reports in June 2026. Its announcement lists impressions, pages, countries, devices for Search, and dates. An August update says the insights reached websites worldwide. These impressions describe appearances of site URLs within Google’s generative features. They should remain distinct from visits and from a panel’s sampled answers. Google’s reporting announcement.
Bing’s AI Performance reporting includes citation counts, cited pages and sampled grounding queries, the phrases used to retrieve supporting information. Its June 2026 Citation Share feature measures a site’s citations against all citations for the same grounding query. Microsoft describes that measure as observational and states it does not represent traffic share. Bing’s initial reporting definitions, June reporting update.
For each export, save the property, date range, filters and fields. If a report cannot be opened or is empty, say which condition applies. Avoid joining unlike platform percentages into a single score.
Track visits with consistent source definitions
Check actual incoming traffic before creating an AI source group. OpenAI states that ChatGPT referral URLs include utm_source=chatgpt.com. Validate that the parameter survives the site’s redirects and appears in analytics. OpenAI’s publisher FAQ.
Keep observed referral domains and campaign parameters in the evidence record. Document the rules used to combine them. If a rule changes, recompute the earlier comparison where possible or mark the break.
In Google Analytics, first-user dimensions describe acquisition of the user, while session dimensions describe acquisition of individual sessions. Event attribution provides another view. Choose the scope required for the question and label it in the report. Google’s traffic-source scope definitions.
An illustrative buyer may arrive from an AI answer, return through a search result and submit a form later. The source associated with the submitting session can differ from the source that first introduced the buyer. A source report therefore needs a stated attribution rule.
Ask how the buyer heard about the business when appropriate. Keep that self-reported answer alongside observed source data. Neither should overwrite the other. Consent choices, device changes and unrecorded visits can leave gaps in the observable path.
Connect enquiries to the reporting period
A click on a submit button is insufficient evidence of a completed enquiry. Validate the successful-submission record, remove duplicates and join it to qualification where a legitimate record link exists.
Use an enquiry reference rather than personal details in the reporting worksheet. Record submitted date, landing page where known, observed source, self-reported source, qualification status and review date.
Allow unresolved enquiries to remain unresolved. If three requests arrived yesterday and sales has reviewed one, present that status. Treating the other two as unqualified would confuse incomplete review with poor lead quality.
Compare groups of enquiries from the same period and track their age. A recent month’s leads have had less time to become opportunities than leads from the previous quarter. Define a review window that fits the sales process and recheck earlier groups when that window closes.
This connects search and AI visibility to conversion measurement without assuming that every citation generated a lead.
Use this monthly reporting worksheet
Create one worksheet with the following five records. Assign an owner to maintain it and another person to review commercial outcomes.
1. Measurement specification
Record the reporting dates and timezone, panel version, target markets, products, question sources, repetition schedule and exclusion rules. Add the names of analytics properties and the date each export was taken.
2. Answer evidence
Use one row per question, product, market and repetition. Include run status, saved answer, domain citation present, linked page, brand mention present, description accuracy and reviewer notes. Keep failures visible. Preserve an audit trail when a reviewer corrects a classification.
3. Website evidence
Record source, attribution scope, landing page, sessions and completed enquiries for the same reporting period. Attach the source-grouping rules. Mark unavailable dimensions explicitly so a blank cell cannot be mistaken for zero.
4. Commercial review
Record enquiry reference, received date, qualification status, reason, owner and next review date. Count unique enquiries. Include opportunity and revenue fields only when the organisation has verified those outcomes and defined its attribution method.
5. Decision record
For each proposed action, write the observation, competing explanations, change proposed, owner and check date. Identify the metric expected to respond and the evidence that would justify continuing or stopping the work.
For example, a supplier-comparison page repeatedly cited by competitors may cover procurement questions absent from your equivalent page. Verify that gap and update the existing page if it serves the same decision. Record the change date and review the same panel afterwards. A subsequent improvement supports further investigation; the before/after comparison alone cannot isolate its cause.
Frequently asked questions
How many prompts are enough?
The number depends on the buyer decisions, markets and products the team needs to cover. Start with a set that can be sourced, checked and repeated reliably. Explain what it covers and where it has gaps. A larger list of near-identical questions can increase the apparent sample without adding useful coverage.
What if citations increase but enquiries do not?
Inspect which questions gained citations, which pages received visits and whether those visits reached a relevant next step. Check enquiry measurement and qualification delays. The answer may be weak commercial relevance, a website issue, a long sales process or too little evidence. Choose the next check from the observed gap.
Can a content change prove an AI visibility gain?
A before/after comparison can show that visibility changed after an edit. Other content, model behaviour and user demand can also change. Preserve repeated observations and comparable unchanged pages where feasible. Describe the result at the level the study supports, with its dates and limitations.
Does the site need special AI markup?
Google states that its AI features require no special schema or new machine-readable file. Eligibility still depends on search requirements, and inclusion is not guaranteed. Check existing access, indexing and visible content before commissioning technical additions. Google’s AI feature guidance.
For a first AEO measurement review, bring a sourced question list, a defined website conversion and the latest qualification record. Those inputs allow the review to end with a specific page decision and a date to check its effect.



