Why do Bing's AI citation counts differ from my tracking tool?

Microsoft's AI Performance report in Bing Webmaster Tools counts citations from real usage. It covers Microsoft Copilot, AI-generated summaries in Bing and select partner integrations. A prompt tracker counts citations only in the answers to its own prompts, run on its own schedule.

So the two numbers answer different questions:

• Microsoft's report: how often did Microsoft's AI surfaces cite my pages across everything users asked?
• Your tracker: how often did each engine cite me for the specific prompts in my prompt set?

The first is close to a full count for one engine family. The second is a sample across several engines. A large gap is the expected result, not a bug. The fix is not to reconcile them into one figure. It's to use each for the job it can do, which the comparison of prompt tracking and engine-reported data covers in more depth.

OtterlyAI ran a public test in 2026: two new domains, each loaded with roughly 1,000 AI-generated posts, tracked across seven AI engines for about four months.

For capybara-world.com, Otterly's prompt-based tracker logged 3,934 citations in total. Of those, 2,718 came from Microsoft Copilot. Microsoft's own AI Performance report credited the same site with 42,817 citations.

Two things matter when you read those numbers:

• Compare like with like. The 3,934 figure spans seven engines. The fair comparison with Microsoft's report is the Copilot slice, 2,718 against 42,817.
• The windows may differ. Otterly's published figures and Microsoft's report may not cover the same dates. Don't read the gap as an exact ratio.

Even with those caveats, the direction is clear. Where an engine publishes its own count, that count sees more than a prompt sample can. The test also has limits: one site, a low-competition topic, and a vendor running it. It shows what a sample is, not how big the gap will be for you.

A tracker is limited by design. Several mechanisms push its count below the engine's:

• Real users phrase things in endless ways. A tracker asks a fixed list. Every variation it doesn't ask is a citation it can't see.
• Engines rewrite the question. OpenAI says ChatGPT search typically rewrites a query into one or more targeted queries before sending them to search providers. The query that retrieves your page may look nothing like your prompt.
• Retrieval is discretionary. On every major platform, the engine decides per query whether to search at all. A prompt that triggered a search yesterday may not today.
• Trackers run at intervals. A daily or weekly run catches a snapshot. The engine logs every answer in between.
• Some surfaces are out of reach. Microsoft's report includes partner integrations that an outside tracker can't query.

This is why adding more prompts raises raw citation counts without telling you the true total. A bigger sample is still a sample.

Each source wins on different questions.

Rule of thumb: trust the engine for volume on its own surfaces. Trust the tracker for content, accuracy and side-by-side comparison. Bing's grounding queries are especially useful because they show what the AI actually searched for. You can see which queries Copilot used and feed the best ones back into your prompt set.

Neither count is perfect. Watch for these:

• Grounding queries are sampled. Microsoft scopes its sampling note to the grounding queries metric, not to the whole report. The details are in whether Bing's AI Performance data is sampled.
• Citation Share is not a scoreboard. Microsoft calls it an observational metric. It can fall while your absolute citations rise, because the denominator grows with competitors.
• The report is new. It entered public preview in February 2026 and expanded in June 2026. Definitions can change, so note the date on every export.
• Cited is not the same as used. OpenAI exposes retrieved sources separately from citations, and the retrieved set is larger. Mechanistic research on how models cite shows citation output can be changed without changing the sources. Both counts are proxies for influence.
• Google shows less. As of October 2026, Google's Search Console report for AI Overviews and AI Mode shows impressions only, with no clicks or position. Google also says the two surfaces may use different models.
• Trackers have partial runs. If an engine errors on some prompts, the count drops for reasons unrelated to you.

1. Verify your site in Bing Webmaster Tools and open the AI Performance report. The Bing Webmaster Tools setup playbook walks through it.
2. Export on a fixed schedule. As of our August 2026 evidence, the report had no API, so manual exports are the only route.
3. Match windows before comparing. Line up the exact dates in both sources, and compare Microsoft's total with your tracker's Copilot-only count.
4. Mine the grounding queries. Add real retrieval phrasing to your prompt set so your sample looks more like actual traffic.
5. Trend each source separately. Never average an engine count with a tracker count. They have no shared denominator.
6. Use a tracker for what answers say. For a quick snapshot, the free checker on this site shows whether engines name you for up to three buyer questions. It's a sample too, so read it alongside Microsoft's report.

If Copilot citations are low in both sources, check access first: whether Bing has indexed your pages and whether your CDN lets bingbot through. How Copilot grounds answers in Bing explains why the Bing index comes first.