What the preview actually reports
No screenshots here, deliberately: the interface is a preview and will drift. Described honestly, the report is a dashboard inside Bing Webmaster Tools for a verified property, with a selectable date range, four measurement blocks, and a trend graph.
Total Citations counts how many times your pages were displayed as sources in AI-generated answers during the selected period. It is an aggregate event count across all users of the supported surfaces, which is exactly what your sampled panel can never see, and why the two must stay separate.
Average Cited Pages reports the average number of unique pages from your site displayed as sources per day. Together with total citations it tells you whether citation activity is concentrated on one page or spread across many.
Grounding Queries lists key phrases the AI used when retrieving content that ended up cited. Two cautions: Bing says this is a sample, and these are retrieval phrases, not the text users typed. Treat them as themes, never as a prompt-demand dataset.
Page-level Citation Activity shows citation counts for specific URLs, revealing which pages participate in AI answers at all. The trend timeline puts the citation counts on a time axis, which is where over-interpretation usually starts.
How the data flows, and why it is not a rank
The diagram below traces both pipelines. In text: Bing’s pipeline runs from your site pages (subject to robots.txt), through Microsoft’s retrieval, into AI answers that display sources, and finally into the aggregated AI Performance report. Your pipeline runs from a frozen prompt panel into an observation log. Both feed the client report as separately labeled sources with different denominators, and they are never summed.
The missing arrow is the point: nothing in either pipeline reports where a citation appeared inside an answer, how it was presented, or whether anyone acted on it. Bing states this directly, which is more honesty than most third-party AI visibility dashboards manage. A rising citation trend is evidence of participation, not position.
Client-safe wording for each metric
| Metric | Supports saying | Does not support saying |
|---|---|---|
| Total Citations | “Bing reported N citations across supported AI experiences this period” | “We rank highly in AI answers” |
| Average Cited Pages | “An average of N unique pages per day appeared as sources” | “Our whole site is AI-optimized” |
| Grounding Queries | “Sampled retrieval phrases clustered around these themes” | “Users are asking exactly these questions at this volume” |
| Page-level Citation Activity | “These URLs were cited most often in the report” | “These are our most important or best-ranked pages” |
| Trend timeline | “Reported citation activity rose after [date]; cause unestablished” | “Our changes produced these citations” |
Folding it into the measurement protocol
The preview slots into the measurement protocol as a first-party evidence column with three rules.
Log every export. Preview interfaces change, and a trend across changed definitions is broken. Record the date range, filters, the interface’s current metric definitions, and known site changes during the period, every time.
Bing AI Performance export logOne row per export: property, date range, comparison period, the four metric values, top cited URLs, grounding-query themes, current interface definitions, site changes in the period, and the panel wave it is reported beside.
CSV templateKeep denominators apart. Bing counts citation events across all users of its surfaces; your panel counts your own frozen observations. A workable report sentence: “Bing Webmaster Tools reported 84 citations to 11 pages during July. In our separate 30-prompt, three-run Copilot panel, an owned page was cited in 14 of 90 valid observations. These datasets have different coverage and are reported separately.” Fold both into their own labeled subsection of the client report, and never one chart.
Use each source for what it is good at. The panel answers “what happens on the buyer questions we chose”; the Bing report answers “how much citation activity exists beyond our sample.” Disagreement between them is information: heavy reported citation activity alongside a silent panel usually means your panel is missing question territory where the site already participates.
From cited pages to next actions
Citation data earns its place when it changes work priorities. Check the earliest layer of the four-layer model before celebrating or panicking, then use the table.
| Situation in the report | Check first | Likely action |
|---|---|---|
| Priority page cited often | Representation accuracy of that page’s facts in sampled answers | Fix any stale or conflicting fact before amplifying the page |
| Priority page never cited | Eligibility: response, robots, index controls, rendered text | Clear the earliest technical blocker, then re-observe |
| Wrong page for the intent cited | Canonical and duplication across the candidate URLs | Consolidate so one preferred URL carries the intent |
| Grounding themes off target | Whether evidence exists for the themes you actually want | Build or strengthen the missing evidence pages |
| Sudden trend drop | Interface definition changes, then site changes in the period | Investigate with dates before reporting a loss |
After a material content fix, Bing positions IndexNow as the way to notify participating engines of changed URLs, per its documentation; the honest framing is “we notified, then monitored,” never “we resubmitted, so citations will follow.” Nothing on this page or in the Bing report guarantees a future citation.
Keeping this denominator separate is a scored checkPlatform counts and panel observations are different populations, and the protocol grades whether your record keeps them apart, alongside 46 other checks on scope, eligibility, sampling and handoff.
Score the evidence record