Exporting and Analyzing Bing Grounding Queries

Most AI visibility tools test prompts that someone chose in advance. The AI systems don't search with those prompts. OpenAI says ChatGPT search typically rewrites your query into one or more targeted queries before it sends them to search providers. Google's Gemini API grounding automatically generates one or more search queries and runs them. The real retrieval input is invisible on almost every platform.

Bing is the exception. The AI Performance report in Bing Webmaster Tools, in public preview since February 2026 and expanded in June 2026, lists grounding queries: in Microsoft's words, the key phrases the AI used when retrieving content that was referenced in AI-generated answers. Coverage spans Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations.

That makes this the only first-party record of what an AI actually searched for before citing you. It tells you which topics already pull you into answers and where the AI's phrasing differs from yours. It doesn't tell you what ChatGPT, Gemini or Perplexity ran. Treat it as Microsoft-only evidence. If you want the background on how that pipeline works, read how Copilot grounds answers in Bing.

You need a verified site in Bing Webmaster Tools with the AI Performance report visible. If you haven't set it up, weigh the effort first: is Bing Webmaster Tools worth setting up for AI search covers the trade-off.

Three facts shape how you read the data:

• There is no API, as of the evidence date. The Bing Webmaster API reference had no fields for citations or grounding queries in the August 2026 review. Microsoft has indicated API access during 2026, so check the current state at whether there is an API for Bing's AI Performance report. Until then, the export is manual.
• Grounding queries are sampled. Microsoft says the grounding query data represents a sample of overall citation activity. That caveat sits on grounding queries specifically, not on the whole report. More detail is in whether Bing's AI Performance data is complete or sampled.
• Only cited retrievals show up. The definition covers content that was referenced in an answer. Pages that were retrieved and then dropped never appear.

One more calibration point. In Otterly's 2026 experiment, Microsoft's own 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 isn't a clean ratio. It does show that an engine's own count sees more than any prompt sample.

Run this once a month. It takes under an hour for most sites.

1. Open the AI Performance report in Bing Webmaster Tools for the verified property.
2. Set the date range to the last full calendar month. Write the exact start and end dates down. Sampled data from different windows can't be compared fairly.
3. Open the grounding queries view. If the report offers an export option for that table, use it. If it doesn't, copy the visible rows into a spreadsheet, and take dated screenshots as a backup.
4. Capture the context columns. Record the cited page for each query where the report shows it. Record the report's Intent and Topic labels where they appear. Microsoft's intent labels include Informational, Commercial, Navigational, Learn and Solve, Research, Creation and Local.
5. Capture the headline numbers for the same window: Total Citations, Average Cited Pages and Citation Share. Keep them on a separate tab.
6. Name the file consistently, for example bing-grounding-2026-09.csv, and store each month in one folder.
7. Add a notes column for anything that changed that month: a site migration, a pricing page rewrite, a new product.

Keep the raw export untouched. Do all grouping in a copy, so you can redo the analysis if your categories change.

Raw grounding queries are long, varied and repetitive. Grouping turns them into decisions.

Topic

1. Read the full list once without sorting.
2. Write 5 to 12 topic labels that describe what you sell and what you know. Use the report's Topic labels as a starting point if they fit.
3. Tag every query with one topic. Add an "off-topic" label for queries that have nothing to do with you.
4. Count queries per topic.

Intent

1. Use the report's Intent label where it exists.
2. Where it doesn't, assign one yourself with the same vocabulary, so months stay comparable.
3. Make a simple pivot: topic down the side, intent across the top.

Map to pages

For each topic and intent cell, write down the page that should answer it. Then check whether the cited page matches. Say you run a hypothetical accounting software company. The table might look like this:

Row one is a mismatch: the AI is quoting an old post on a commercial query. Row two is healthy. Row three is a gap. When you see queries you never targeted, why Copilot cites pages for queries you never targeted explains the usual causes.

Good signs

• Most queries fall into topics you actually sell or know well.
• The cited page is the one you would choose for that query.
• Commercial queries land on pages that state current prices and plans as plain text.
• The same topics recur month to month, even if exact wording changes.

Warning signs

• Old or wrong pages cited on commercial queries. The AI may repeat outdated facts from them. Fix the facts on the source page and make sure the current page states them clearly. A specific fact on a page that can be fetched gives the engine a sentence it can quote. That is a chance, not a promise.
• Many off-topic queries. Your content is being pulled into answers you don't want to be part of. Decide whether those pages still serve you.
• A topic you sell with zero queries. Check that the relevant page is in Bing's index and that its key content is in the raw HTML.
• Zero rows across the board. This may mean you're not being cited. It may also mean crawlers can't reach you. Check access separately, for example by reading server logs for AI crawler status codes.

Don't read a falling Citation Share as failure on its own. Microsoft describes it as an observational metric, not a ranking system or a competitive scoreboard. It can fall while your absolute citations rise.

• Same date range logic as last month (one full calendar month).
• Raw export saved untouched, with exact dates in the filename.
• Cited page recorded for every query where the report shows it.
• Intent and Topic labels captured, or assigned with the same vocabulary as before.
• Headline numbers (Total Citations, Average Cited Pages, Citation Share) logged.
• Every query tagged with one topic, including off-topic.
• Topic by intent pivot built.
• Each cell mapped to the page that should answer it.
• Mismatches and gaps listed as fixes with an owner.
• Site changes from the month noted.
• No conclusions drawn from a single sampled month.

Monthly is the right default. It matches how the report is easiest to window, and it gives sampled data time to show a pattern.

Add an extra run after a change that could affect retrieval: a site migration, a rendering change, a robots.txt or CDN update, or a rewrite of pricing or product pages. Compare the next full month against the month before the change.

Look at trends over three months or more before acting on a small topic. A topic that appears in one month and vanishes the next is often sampling noise.

Remember what this data can't do. It covers Microsoft's surfaces only, and it shows queries tied to citations, not every retrieval. For the other engines, pair it with a direct check of what they say about you. The free checker on this site shows whether ChatGPT and other engines name you for up to three buyer questions, as a one-time snapshot.