In short
- Give every metric a defined population and reporting period before comparing results.
- Use platform reports alongside a fixed question panel, with branded questions reported separately.
- Keep identified visits and qualified enquiries separate from observations inside AI answers.
Start with the decision you need to make
Before opening a dashboard, write down the question your report should answer. Perhaps buyers cannot find your product when they describe a particular problem. Perhaps an assistant quotes an outdated price. Those problems need different observations.
For discovery, inspect answers to questions that don't name your company. For accuracy, ask about your company and compare the answer with current product information. For commercial impact, examine identified visits and subsequent enquiries.
Use explicit counting rules. In your review, a mention means the answer names your brand. A suitable recommendation means it proposes your product for the buyer's stated need. A citation means a visible source link. Keep a separate field for an incorrect product claim.
These categories can overlap. Don't add them into a combined success total. Count an answer once within each category, even when it repeats the brand name.
Define retrieval separately if your tool exposes it. Record what that field actually proves. A source link alone cannot establish every page the system consulted.
Read Bing's AI Performance report within its scope
Bing Webmaster Tools reports citations across Microsoft Copilot, AI summaries in Bing and selected partner integrations. It identifies cited pages and shows activity over time. 1
Its grounding queries describe phrases used to retrieve cited content. Microsoft describes that query data as a sample. These phrases aren't presented as a complete transcript of users' conversations. 1
Start with the pages relevant to your business question. Save the date range alongside the export. Check whether citations concern a product explanation or an unrelated article before deciding what to change.
Microsoft also documents Citation Share for a grounding query. Its denominator is citations across sites for that query. It doesn't represent traffic share or a ranking. 2
A citation count doesn't tell you whether the answer recommended your company. The launch documentation also excludes conclusions about placement within an answer. 1 Keep those limits in the report caption.
Don't label the aggregate as your visibility across every assistant. If a surface cannot be isolated, retain the platform's combined scope. Use a separate answer sample to investigate brand recommendations.
Use Search Console's dedicated AI report
Google's Generative AI performance report for Search records impressions from AI Overviews and AI Mode. Discover has a separate report. Search Labs experiments are excluded. 3
Read the metric literally. An impression records a displayed link under Google's counting rules. The documented dimensions include pages and countries, with device and date breakdowns. They don't provide the full answer text. 3
Export comparable periods for the pages you changed. Keep search type and filters consistent. Label these observations as Google Search AI impressions. Avoid turning them into a brand recommendation rate.
The AI data also belongs to the Web performance reporting scope. Don't add both reports together as independent exposure. Property totals and page totals can differ because their aggregation differs. 3
An absent report doesn't establish zero visibility. Google says insufficient impressions can prevent it appearing. Its documentation also identifies exclusion settings as a possible reason. 3
For branded versus unbranded answer analysis, use your labelled question panel. Don't infer that split from a page URL alone. A pricing page could appear in either kind of answer.
Follow identified AI visits in GA4
GA4's Traffic acquisition report uses session dimensions to describe incoming traffic. Start with Session source / medium and keep the chosen reporting period visible. 4
The default AI Assistant channel groups visits from recognised assistant sources. Google explicitly excludes AI Overviews and AI Mode from that channel. Organic clicks from them are classified as Organic Search. 5
Use the channel as a starting view, then inspect its source rows. Record any custom classification separately. A rule changed midway through a comparison belongs in the report notes.
Review landing pages and your configured key events alongside those visits. GA4 includes sessions and key events in Traffic acquisition. 4 Check that the event represents the action you intend to count.
A form opening is different from a submitted enquiry. A submitted enquiry still needs qualification. Define those stages before anyone presents an acquisition result to the sales team.
GA4 classifies traffic without a clear referral source as direct. Missing information and redirects are among the documented explanations. 6 You cannot assign that unidentified traffic to AI from the label alone.
Describe the result as identified AI referrals. It doesn't measure everyone who encountered your brand in an answer. Avoid multiplying the total by an assumed correction factor.
For a fuller investigation, record customers' voluntary descriptions of how they found you. Keep that information separate from session attribution. If both describe the same enquiry, avoid counting it twice.
Build a question panel you can keep
A tracker such as Peec AI adds a controlled set of questions to these platform observations. Peec's methodological guidance recommends stable comparisons and separate branded reporting. 7
Choose questions from actual buying conversations where you have permission to use them. Remove confidential details. Include product constraints that affect the decision, such as deployment requirements or an integration the buyer needs.
A fictional facilities team might ask which scheduling products work without reliable connectivity. A branded question might ask whether a named product supports that workflow. Keep these in separate groups.
For branded questions, assess the accuracy of the answer and its suitability for the task. Simply repeating a name already supplied in the question is weak evidence of discovery.
Write down the exact prompt and assign a panel version. Record the engine that produced the answer, even when the question mentions a different engine. Include the market and language.
Also record the product surface used and any known model setting. Ask the tracker how it collects answers. Don't assume an API response reproduces a consumer session without checking its collection method.
Keep the original panel when adding new questions. Report the extension separately until you have a comparable baseline. Peec specifically recommends comparing questions tracked throughout the entire period. 7
Store failed runs with their reason. Decide exclusions in advance and show the completed-answer count. A run that failed technically must remain distinguishable from a valid answer that omitted your brand.
Report repeated observations and their range
A screenshot documents an answer worth investigating. It doesn't establish how often that answer occurs. Peec's analysis distinguishes variation across repeated runs from differences caused by question selection. 7
Our recommendation is to summarise repeated observations across comparable reporting windows. Show the number of completed answers beside the count naming your brand. Preserve the underlying answers so another reviewer can inspect them.
Keep each engine separate. Within an engine, separate branded questions from discovery questions. Show question-level results when an aggregate changes enough to affect a decision.
For an observed range, report the lowest and highest rates across the same predefined subperiods. Include the denominator for each subperiod. Call this an observed range. Don't label it a confidence interval without a justified statistical method.
Report the aggregate rate alongside that range. Calculate it from the pooled counts, or disclose any weighting that gives each prompt equal influence. Avoid silently averaging percentages with different denominators.
Repeated runs improve your view of the chosen questions. They don't make your panel representative of every buyer. Treat a small panel as exploratory. It shows where to look, without supporting a claim about overall visibility. Keep that sampling limit visible when presenting results outside the team.
Connect the report to the next investigation
Suppose citations rise after a product page changes, while suitable recommendations remain flat. Inspect the answers and the cited passages before drawing a conclusion. The page might be supporting a general explanation.
If identified visits rise, check which pages received them and whether the enquiries fit your business. Keep that finding distinct from any claim about the edited page causing the increase.
Maintain a dated change log beside the reporting exports. Record concurrent site changes and known measurement changes. Where practical, compare similar unchanged pages, while acknowledging that a comparison alone doesn't establish causation.
Use the AI visibility review sheet to write down your panel and counting rules. Add the underlying answer for any finding that drives a decision.
End each review with a specific next check. That could be verifying a price, fixing an inaccessible page or investigating an irrelevant citation. A useful report leaves someone knowing what to inspect next.
Sources
- Microsoft Bing, Introducing AI Performance in Bing Webmaster Tools Public PreviewAccessed Back to the text, first mention of source 1
- Microsoft Bing, New AI Visibility Insights in Bing Webmaster ToolsAccessed Back to the text, first mention of source 2
- Google Search Console Help, Generative AI performance report (Search)Accessed Back to the text, first mention of source 3
- Google Analytics Help, Traffic acquisition reportAccessed Back to the text, first mention of source 4
- Google Analytics Help, Default channel groupAccessed Back to the text, first mention of source 5
- Google Analytics Help, Understand (direct) / (none) trafficAccessed Back to the text, first mention of source 6
- Peec AI, The key to prompt tracking (vendor's measurement research)Accessed Back to the text, first mention of source 7
