How to see when AI cites you, using Bing's own data

Start with scope, because it decides what the numbers can mean. Microsoft says AI Performance reflects citation activity across three kinds of surface:

• Microsoft Copilot, the assistant.
• AI-generated summaries in Bing, the generated answers on Bing's own results pages.
• Select partner AI integrations, third-party experiences built on Bing. Microsoft does not publish which partners these are, so do not read the report as a ChatGPT, Perplexity or Gemini dashboard.

The unit is the citation: one instance of your page being visibly referenced or shown as a source in an AI-generated answer. It is not a ranking, it is not a measure of how important the page was to the answer, and by Microsoft's own FAQ it is not a click, a visit, or any kind of engagement.

Follow one question through the system and it leaves three records, each in a different tool:

Only the middle line is new. What makes it valuable is how close it sits to the decision you care about: whether the engine chose your page as a source to show the person asking.

AI Performance lives inside Bing Webmaster Tools, so the only prerequisite is a verified site. If you already use Google Search Console, the import path verifies you without touching DNS.

1. Sign in to bing.com/webmasters.
2. Add your site. The fastest route is Import from Google Search Console, which reads your verified properties and sitemaps and verifies the sites you pick automatically. Otherwise verify with an XML file, a meta tag, or a DNS CNAME record.
3. Confirm your sitemap is submitted. The report only reflects content eligible for Bing's index, so a page Bing has not found cannot be cited in its answers.
4. Open AI Performance (direct link: bing.com/webmasters/aiperformance).
5. Pick a range: 7 days, 30 days, 3 months, or a custom range within the history the dashboard holds.

The top of the report is a handful of counts and a timeline. Each is simple on its own. The insight comes from reading them against each other.

Volume versus breadth

Read total citations and average cited pages together. If citations climb while average cited pages stays flat, more of your citations are landing on the same few pages, so a large part of your visibility rides on those pages staying accurate and current. If average cited pages rises, citations are spreading across more of your site, and the total depends less on any single page.

The timeline shows what, never why

Microsoft names three kinds of cause behind a moving line: the volume and type of questions people ask, updates to your content, and system or model changes on its side. It is explicit that a trend cannot be pinned on any single one of them. Treat a rise after a content update as consistent with your change, not proof of it, and read Part 07 before you celebrate a sudden jump.

In the AI answers this report covers, grounding is the retrieval step that brings current web content into the answer: the system can run searches, pull in relevant pages, and cite sources it used. A grounding query is Microsoft's summary of the phrase used in one of those searches, when it retrieved content that ended up cited. It is a rare look at the engine's side of the conversation, with one catch that trips up nearly everyone.

A grounding query is not what the person asked. Microsoft describes the phrases as grouped and generalized. The report does not show individual prompts or specific long-tail questions, and different AI surfaces and partners may phrase them differently. One question can fan out into several retrievals, and many different questions can collapse into the same short phrase.

what the person asks Copilot
  "we're a 12-person design agency, what should we use to manage client work?"

what the report shows
  project management for agencies .......... 38 citations

the page it maps to
  /compare/agency-project-tools

The number beside each phrase is how often content from your site was referenced in AI answers associated with it during your range. One phrase can map to several of your pages, and one page can sit under several phrases.

Follow the mapping in both directions

Since March 2026 the report connects the two views. Select a grounding query to see which of your pages were cited for it, or select a page to see which phrases led to it. You can filter by one or the other, not both at once, and an export reflects whichever filter is applied, so export filtered views when you need to cross-reference.

Three patterns are worth hunting for:

• Where you are already cited. High-citation grounding queries mark the subjects where your content is already being used as a source. Those are the pages to keep current first.
• The wrong page answering. If your pricing page is cited for "how does X work," you may have no page that answers that question directly. Writing one gives the engine a better candidate, though whether it switches is the engine's call.
• The phrase that is missing. If a core subject of your business never appears, you are either not cited for it or cited too rarely to surface. Either way, that is where to look next.

On June 16, 2026, Microsoft added four views to the report, all still labelled preview. Together they move it from "how often" toward "in what context" and "compared with what."

Intents: what kind of question you are cited for

A machine-learning classifier gives each grounding query an intent label, from a taxonomy that goes well past the classic informational, navigational and transactional split:

• Informational
• Media
• Navigational
• Commercial
• Learn and Solve
• Research
• Live Event
• Local
• Comparison
• Planning
• Utility
• Creation
• Conversational
• Other

The useful read is the mix. A software company cited almost entirely on Informational phrases is showing up when people learn about its category, not when they compare options or get close to buying. Microsoft suggests matching format to intent: comparison tables for Comparison, step-by-step structure for Planning, concise explanations for Informational. It also warns that labels can misfire on ambiguous or multi-intent phrases, so judge the distribution, not a single row.

Topics: the themes above the phrases

Topics cluster related grounding queries into themes, so "solar panels" and "residential solar installation" roll up into something like Solar Energy. This is the view to bring to a content planning meeting, because teams plan in themes, not phrases. Two caveats: labels can be broad in niche or specialized fields while the classifier matures, and seeing which pages are cited for a topic currently means exporting the data and filtering it yourself.

Citation Share: your share of the sources

Citation Share is the first relative number in the report, and it corrects a trap in raw counts: a busy phrase can hand you plenty of citations while you are still one voice among many.

  your citations for the query
÷ all citations shown for that query, across every site
= Citation Share

Know what it is not. Citation Share does not name the other domains, it is not a share of traffic, and it is not a quality score. Microsoft describes it as observational, not a ranking or a competitive scoreboard.

Compare: this period against the last

Compare overlays an earlier period on the chart: the current range as a solid line, the comparison as a dashed one, aligned day by day. Use the previous period or pick two custom ranges. It is the right tool for checking a page 30 days after you rewrite it, as long as you remember it shows that something changed, not why.

Bing Webmaster Tools reports citations. Microsoft's free analytics product, Clarity, now has an AI Visibility section that covers what sits either side of them: the bots that read your pages and the visitors who arrive. Microsoft describes Clarity's Citations dashboard as unifying signals from Bing Webmaster Tools and Clarity, so the two are best read side by side.

Share of authority reads high by design

Clarity's share of authority is a cousin of Bing's Citation Share: your citations divided by total citations from all domains. The difference is in what gets counted. Clarity calculates it day by day and includes a query only on the days your domain was cited for it, so the days you were absent never pull the number down. Microsoft notes that this can produce higher values than a broader calculation would. Read it as your standing where you already appear, not your share of a whole topic, and do not expect it to match Bing's figure.

If you read early coverage of this dashboard, you may have seen query volume and citation rate metrics. Microsoft's February early-access announcement described them, but the current documentation, updated after Citations became generally available, no longer lists either, so do not plan around them.

Topic Insights is a simulation, and says so

Topic Insights works differently from everything above. Instead of reporting citations from real usage, it runs the prompts you write through a model (Microsoft names GPT-5.3) grounded by Web IQ, Microsoft's search layer for AI systems, then analyzes which domains get cited and how much of each answer they support. That makes it a practical way to test the questions you care about and see competitors by name. Microsoft is equally clear that it is for directional monitoring only. It is not a reading of what ChatGPT, or any particular assistant, tells real users. AI Visibility is also limited to eligible website categories, so some projects will not see every report.

AI referrals, with the session attached

The Citations dashboard puts AI referral traffic beside the citation counts: sessions that arrived from AI assistants as a percentage of all your sessions. Behind that number, Clarity sorts visits from chat assistants such as ChatGPT, Claude, Gemini and Perplexity into an AIPlatform channel and ties them to its recordings and heatmaps, so you can watch what AI-referred visitors actually do. Like GA4, it cannot split AI features that share a search engine's referrer from ordinary search clicks. One practical note: Clarity is a tag on your pages, so it falls under the same consent rules as any other analytics cookie.

Every limitation below comes from Microsoft's own documentation. None of them makes the report useless. All of them change how you should read it.

• It is a sample. The data is aggregated to be representative, not a log of every citation, and very low or infrequent activity may not appear at all.
• Totals will not reconcile. The pages view, the grounding queries view and the timeline can show different totals. Filtering by a phrase and then a page can give a different count than the reverse, because each is sampled over slightly different windows inside your range.
• Numbers can move after the fact. Results may be refined as more data is processed.
• History is short. You get only the ranges the dashboard offers. Exports come as CSV or Excel. At launch, Microsoft said the data was not yet in the Webmaster Tools API and that adding it was on its backlog, so check the current state before you build a pipeline, and keep your own monthly exports either way.

The June backfill, or why not to celebrate early

On June 1, 2026, site owners across many accounts watched their citation counts surge. It looked like a visibility win. It was the pipeline: Krishna Madhavan of Microsoft Bing described it as regular data backfilling, with no anomalies. The lesson generalizes. A step change that lines up with a product update or a date boundary is more likely a data event than a real shift, so keep a dated log of your own content changes and of Microsoft's announcements beside your exports.

Put plainly, the report supports some claims far better than others:

Microsoft handles AI use page by page, through the robots meta tags you already use for search, rather than through a separate AI crawler you can block in robots.txt. Bing's Webmaster Guidelines draw the line plainly: robots.txt controls crawl access, not indexing. Blocking Bingbot is on their list of things to avoid: it stops Bing reading your content for search and AI answers alike, and it does not reliably remove a URL. When a URL should not appear in Bing search, Copilot or grounding results at all, the guidelines point to noindex. To stay in search while limiting AI use, reach for the tags below.

A few details matter. Each tag also works as an X-Robots-Tag response header. Writing name="bingbot" instead of name="robots" limits a tag to Bing. A page carrying both nocache and noarchive is treated as nocache, and the guidelines warn that nocache reduces citation depth, so they advise against it on content you want used in Copilot. Finally, nosnippet and data-nosnippet stop Bing showing captions and may limit Copilot citation quality. The finer tool is the data-snippet attribute, which marks the text Bing may display or cite.

The February 2026 guideline rewrite also added a section on prompt injection and AI manipulation. Content written to steer or interfere with Bing's language models is now a named violation, not a clever trick.

Everything above, turned into a routine you can run in about an hour a month.

• Set up once. Verify in Bing Webmaster Tools (import from Search Console), submit your sitemap, and turn on IndexNow, which Microsoft recommends so content changes are discovered faster.
• Export monthly. Download grounding queries, pages and the time series as CSV or Excel. History in the dashboard is limited, so your archive is the only long-run record.
• Keep a change log. Date every content update and every Microsoft announcement beside the export, so any jump can be checked against both.
• Map phrases to pages. For your top grounding queries, confirm the right page is the one being cited. Where the wrong page answers, write or fix the right one.
• Track share on core phrases. Pick the ten grounding queries closest to revenue and record their Citation Share. Low share on a core phrase is your priority list.
• Read the intent mix. Cited mostly on Informational phrases while you sell something? Check whether you have the comparison and buying content that Commercial and Comparison phrases call for.
• Join citations to clicks. Set cited pages beside their Copilot referrals in GA4. A page cited often but rarely clicked may mean readers get enough from the answer that they do not need to visit, but citation placement, the intent behind the question and stripped referrers can produce the same pattern. Where the first explanation looks likely, give readers a reason to click: a tool, original data, or detail an answer cannot carry.
• Audit your directives. Check that no noarchive or nocache has slipped onto pages you want cited.
• Re-check after 30 days. Use Compare on the pages you changed, and record the result as an observation, not a verdict.

The AI Performance report is a real step forward: an engine telling site owners, in its own numbers, when it used their work. It is also one window onto one ecosystem.

It covers Copilot, Bing's AI summaries and partners Microsoft does not name. It says nothing about ChatGPT, Gemini, Claude or Perplexity. Google now reports its own slice: a generative AI performance report in Search Console, rolled out worldwide on August 31, 2026, counts how often links to your site were shown in AI Overviews and AI Mode, by page, country, device and date, with no query dimension. Two engines, two different units, and the largest assistants still unreported.

More important than coverage is what neither report reads: the answer itself. A citation tells you the engine used your page. It does not tell you whether the answer recommended you, described you accurately, or quoted your definition while naming a competitor as the better choice. A brand can be cited constantly and recommended never.

Here is the funnel from the top of this guide again, with what each layer's tools can and cannot see:

• AI access, measured in your logs and Clarity Bot activity: is an engine reading you?
• Answer visibility, partly measured by Bing AI Performance and Google's report: are you shown as a source? What the answer says about you still takes asking the engines directly.
• AI referral, measured in GA4 and Clarity: does any of it send you people?