Bing AI Performance: How to Audit Citations and Grounding Queries Without Calling Them Rankings

Microsoft announced AI Performance in Bing Webmaster Tools on February 10, 2026 as a consolidated view of references to a verified site's content across Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations. On June 16, 2026, Microsoft added Intents, Topics, Citation Share, and comparison controls globally in preview. A current audit must account for both releases rather than treating the February field set as the finished product.

These fields address a different layer from conventional web search performance. Search clicks and impressions describe result exposure and visits under Bing's search methodology. AI Performance describes source references in supported generated answers. Neither layer replaces analytics or CRM evidence, and the dashboard does not observe what happened after a person reached the site.

Traditional rank language suggests an ordered list: position one, position two, and so on for a defined query, market, device, and time. Bing's citation total is not that. It aggregates supported AI surfaces and counts references displayed as sources. The same page can be cited in different contexts, and the metric does not tell the analyst where the link appeared, how visually prominent it was, or whether the answer relied on it for a central claim.

Calling citation count “AI rank” creates false precision. A page with fewer citations might support a narrow, commercially valuable question. A frequently cited informational page might produce no visits. A change could reflect product coverage, query mix, freshness, or answer generation rather than a content release. The number is an observation to investigate, not a final evaluation of the page.

Use Citation Share for bounded comparisons. If a query shows 6 citations for the site and 24 citations across all domains in the sampled period, the displayed share is 25 percent for that query. That does not mean the page appeared in 25 percent of user answers, earned 25 percent of clicks, or outranked a competing domain. The dashboard does not expose competitor domains, so any competitor-source review requires a separately captured, clearly labeled observation outside this dataset.

Bing says grounding queries are key phrases an AI system used when retrieving content that was referenced in generated answers. The help documentation maps queries to cited pages in both directions: one grounding query can cite several pages, and one page can participate in several grounding queries. The displayed data is a sample of overall citation activity. This is a platform-native retrieval clue, not a complete record of what users asked.

A grounding phrase may be generated by the system rather than copied from the user's conversational wording. It may represent one subtopic within a longer research path. Treat it as evidence that the phrase participated in a sampled retrieval associated with a citation—not as a verbatim customer prompt, search-volume measure, or guaranteed keyword target.

The most useful content change is often corrective rather than expansive. A grounding phrase can reveal that an old page is carrying an important topic, that a comparison lacks current specifications, or that a service article has no path to the relevant offer. Fix the evidence and navigation before adding more pages.

Page-level citation activity identifies URLs worth reviewing, while Intents and Topics help group the retrieval context. Prioritize by decision value: strategically important needs, material changes under matching comparisons, evidence defects, stale facts, or broken user paths. High Citation Share is neither a success requirement nor a reason to rewrite a page with weak business fit. Treat classifier labels as machine-generated aids: Bing warns that Intents and Topics can be imperfect, so sample the underlying queries before changing content strategy.

A low-citation page is not automatically a failure. It may serve a narrow need, lack enough eligible demand, or sit outside the supported AI surfaces. The dashboard alone cannot establish which explanation is correct. Review the page's business role and conventional search evidence before deleting, merging, or rewriting it.

Use the Compare control before assigning meaning to a change. A previous-period or custom-period overlay is useful only when filters, page population, and window length are compatible. Annotate releases, indexing incidents, and Bing product changes; then inspect which queries, topics, and pages explain the movement. Comparison is a diagnostic convenience, not a causal model.

Bing connects AI visibility with freshness and recommends IndexNow when content is added, updated, or removed. IndexNow allows a site to notify participating search engines about changed URLs instead of waiting only for routine discovery. That can shorten the path to recrawling, but it does not guarantee indexing, selection, citation, a Citation Share increase, or immediate replacement of an answer.

Sitemaps and IndexNow serve complementary purposes. A sitemap communicates the intended URL inventory and metadata for broad discovery. IndexNow communicates individual URL changes quickly. Neither repairs duplicate pages, contradictory canonicals, an inaccessible origin, or stale visible content. Bing's duplicate-content guidance recommends consolidating signals so the intended version is easier for search and AI systems to identify.

AI Performance is not a click report. Conventional Bing search performance, server or analytics referral evidence, and business outcomes live in other systems with different coverage. A page and date can be used as investigation keys, but temporal proximity does not prove that a cited-source event produced a particular visit or inquiry.

Keep the reconciliation explicit and aggregate. Preserve the AI Performance export at grounding-query, URL, and period level with its filters, intent/topic labels, and Citation Share. Pull conventional Bing search data under its own query and page definitions. Record Bing or Copilot referrals only when request or analytics evidence identifies them. If a referral is absent or ambiguous, leave it unknown; there is no event-level key that turns a dashboard citation into a specific visit.

Run the audit on a stable cadence rather than reacting to every daily move. Export a complete period, annotate releases and incidents, compare a fixed page set, review sampled grounding phrases, and select a small number of page actions. Keep an untouched copy of the export so another reviewer can reproduce the conclusion.

A strong executive summary can be short: which pages were cited, which sampled needs they supported, which evidence or user-path defects were found, and which changes were made. It should also state that the preview does not reveal placement, authority, complete prompts, clicks from a specific citation, or causal impact.