Is Bing's AI Performance data complete or sampled?

Bing Webmaster Tools' AI Performance report went into public preview on 10 February 2026 and was expanded on 16 June 2026. It covers Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations. It reports Total Citations, Average Cited Pages, page-level citation activity, grounding queries, Intents, Topics, Citation Share and a Compare view.

The sampling note is narrow. Microsoft's sentence, "The data shown represents a sample of overall citation activity," sits inside the grounding queries subsection and refers to "this metric." Total Citations and page-level activity carry different caveats, and none of those mention sampling.

So the accurate summary is not "the report is sampled." It is "the grounding query list is a sample, and the citation counts are Microsoft's tally with their own stated limits." That distinction matters when you report numbers to a board or a client. If you call everything sampled, you undersell the counts. If you call everything complete, you oversell the query list.

For a walkthrough of every panel, see reading Bing's AI Performance report.

Each metric answers a different question. Treat them separately.

Grounding queries deserve extra care because they are unusual. They show what the system actually searched for, not what a user typed. Other engines rewrite prompts too: OpenAI says ChatGPT search typically rewrites a query into one or more targeted queries, and Google's Gemini API documentation says grounding automatically generates one or more search queries. Bing's report is the only place we know of that exposes those internal queries to a site owner.

A sample of the right input beats a full count of the wrong one. Just don't sum the sampled queries and present the result as total demand. The guide to exporting and analyzing grounding queries covers how to work with the list.

Microsoft frames Citation Share explicitly as "an observational metric" and "not a ranking system or a competitive scoreboard." It also says the metric does not expose competitor domains, represent traffic share, or assign quality scores to content.

Three consequences follow:

• It is a share, so the denominator moves. Whoever already wins a query dominates the total. Your share can fall while your absolute citations rise.
• It is not traffic. A citation is not a click, and the metric does not claim to measure visits.
• It does not name rivals. If you want to know whether the report shows competitor citations, the short answer from Microsoft's own description is that it does not expose competitor domains.

The next move: report Citation Share next to Total Citations, never alone, and keep Microsoft's "observational" framing in any deck you share.

Prompt trackers count less than Microsoft does

In OtterlyAI's 2026 experiment, Microsoft's AI Performance report credited one test site with 42,817 citations, while Otterly's prompt-based tracker logged 3,934. The time windows may not match, so it is not an exact ratio. It does show that any prompt set is a sample too, and the engine's own count sees more.

Citation is not the same as use

Mechanistic research on retrieval-augmented generation (arXiv 2606.28358) found citation output can be changed without changing the sources the model drew on. A citation count is a proxy for influence, not a measurement of it.

Coverage is Microsoft's surfaces

The report covers Copilot, Bing AI summaries and select partners. Whether it reflects other assistants is a separate question, covered in does the report include ChatGPT citations.

No API yet

As of the evidence we hold (August 2026), the report has no API, though Microsoft's Fabrice Canel has indicated API access during 2026. Check the current API status before building a pipeline.

1. Verify your site in Bing Webmaster Tools if you haven't. The setup guide covers the steps.
2. Export grounding queries monthly. Without an API, a manual export is the only way to build history.
3. Label every number by type. Mark grounding queries as sampled, citation counts as Microsoft's tally, and Citation Share as observational.
4. Compare grounding queries with the prompts you test. If they barely overlap, your prompt set is measuring something the engine doesn't search for.
5. Check the pages that earn citations. Confirm the facts on them are correct and present in the HTML before JavaScript runs.

None of this controls what Copilot cites. It makes sure you read Microsoft's numbers for what they are, and it shows where Copilot's grounding in Bing is drawing from your site.

Can I add up grounding queries to estimate total AI demand? No. Microsoft says the grounding query data is a sample of overall citation activity. Summing a sample gives you a sample total, not demand. Use the list to spot themes, intents and phrasing the AI uses when retrieving your pages, and compare those patterns over time. For volume, look at Total Citations, which Microsoft does not describe as sampled.

Why does my prompt tracker show far fewer Copilot citations than Bing does? A prompt tracker only counts answers to the prompts it asks, so it sees a slice. In OtterlyAI's experiment, Microsoft's report counted 42,817 citations for one site while Otterly's tracker logged 3,934, though the time windows may not match. Where an engine publishes its own count, that count generally sees more than an outside sample can.

Does a rising Citation Share mean I'm beating competitors? Not necessarily. Microsoft calls Citation Share observational and says it is not a competitive scoreboard and does not expose competitor domains. It is a share whose denominator is shaped by whoever already wins each query. It can rise or fall for reasons unrelated to your own performance, so read it alongside your absolute citation count.

Does a citation in the report mean Copilot used my content? It means Copilot showed your page as a source. Research on how language models cite (arXiv 2606.28358) found citation output is partly separable from which sources actually shaped the answer. Treat citations as a useful but noisy signal of influence. Check the cited pages for accurate, clearly stated facts, since that is what the answer can quote.