Bing Webmaster Tools AI Performance Read Carefully

Microsoft introduced AI Performance in Bing Webmaster Tools as a public preview in February 2026. It aggregates citation activity across Microsoft Copilot, AI-generated summaries in Bing, and selected partner integrations. That scope is broader than a single Bing result type and narrower than “the entire AI web.” Record the product status, property, date range, and available filters before exporting anything.

The report extends Bing’s existing search and chat measurement rather than replacing crawl, index, and conventional performance data. A cited URL still needs a successful public response, consistent canonical signals, useful content, and enough freshness to remain a credible source. Citation reporting reveals an outcome; it does not remove the need to diagnose the page and its discovery path.

Total citations counts source displays in supported AI-generated answers. Average cited pages describes the average number of unique site pages displayed as sources per day. Page-level activity assigns citation counts to URLs, while grounding queries expose a sample of retrieval phrases associated with cited content. None of these fields states the source’s visual placement, authority, recommendation strength, or role inside an individual answer.

Preserve Microsoft’s language in dashboards. Do not rename total citations “AI rankings,” average cited pages “share of voice,” or grounding queries “keywords” without a documented transformation and caveat. The cleanest export retains the original metric, product surface, time zone, filter state, retrieval timestamp, and any normalized page-family label added by the publisher.

Group citations by page family, content age, canonical status, and substantive task. A reference page, breaking update, product page, and long-form procedure have different opportunities to be cited. Look for repeated grounding-query themes, not a single surprising phrase. Compare cited and uncited pages only when they solve similar tasks and share the same technical eligibility.

A cited page may already be doing its job and need protection rather than expansion. An uncited page may lack demand, freshness, crawl access, distinct evidence, or a clear answer. The report cannot choose among those causes by itself. Join it to Bing crawl and index data, the production response, server logs, internal-link evidence, and the page’s source ledger.

When a grounding phrase accurately describes the page’s task, inspect whether the direct answer is easy to locate and whether the supporting claim has a date, source, and boundary. Strengthen missing evidence or navigation. When the phrase exposes an adjacent but genuinely different task, create a new brief only if the publication can add a distinct procedure or artifact. Do not turn every phrase into a thin route.

For freshness-sensitive pages, align content updates, honest lastmod, sitemap output, and optional IndexNow notification. An accepted notification does not prove that the new version was cited. Preserve deployment and submission timestamps, then wait for subsequent crawl and reporting evidence before attributing movement.

A useful weekly view shows total citations, cited pages, citation concentration by family, sampled grounding themes, Bing web and chat clicks, referral sessions, and the count of technically eligible pages. Keep citations and visits separate. Many citations may produce few clicks because the answer satisfies the user; a small citation count may still send valuable visits.

Annotate preview changes and avoid week-over-week conclusions on tiny denominators. The report is valuable because it reveals a previously opaque class of source use. Its credibility depends on resisting the temptation to make it a universal AI visibility score. Measure what Microsoft exposes, explain what it omits, and make editorial changes only when the underlying page evidence supports them.