How we measure what people ask AI about money
When you ask an AI assistant a money question, it runs its own searches before it answers you. Microsoft publishes those searches back to the site owners it draws on, along with how much of each answer came from them. This page documents how we collect that data, what the numbers do and do not mean, and where the dataset stops being reliable.
Exported 2026-08-15, covering a rolling 90-day window ending at that date. These figures describe that window and are not live. Anyone quoting them should quote the date with them.
How the data is collected
- The report comes from Bing Webmaster Tools, under AI Performance, Grounding Queries. It is available to any owner of a verified property, for their own site only.
- We download the full report as CSV. Bing does not offer an API for it, so collection is manual and dated by hand rather than scheduled.
- Each row is loaded into a snapshot table, one row per query per export, keyed on the export date. Shares are stored as basis points rather than decimals so that comparing two exports is exact.
- Figures on this page are computed from the most recent snapshot at request time. Nothing here is typed in by hand, so a stale number cannot survive a new export.
What the shape of the data looks like
Demand is extremely concentrated. The single largest grounding query, “compare checking and savings account options”, accounts for 16.93% of every citation in the capture on its own.
| Intent | Share of citations | Our citation share |
|---|---|---|
| Comparison | 33.37% | 21.17% |
| Informational | 23.84% | 24.27% |
| Research | 13.82% | 28.55% |
| Learn and Solve | 13.32% | 26.19% |
| Commercial | 11.89% | 31.55% |
| Unlabelled | 2.68% | 29.78% |
| Utility | 0.83% | 27.26% |
| Others | 0.22% | 31.71% |
Intent labels are generated by Bing, not coded by us. Rows Bing left blank are grouped as “Unlabelled” rather than folded into a named intent, so the breakdown adds to 100% without overstating any single category.
Limits of this dataset
These are the reasons a figure here could mislead. We would rather state them than have someone discover them after quoting us.
- It is one site and one answer engine. Microsoft Copilot and the partners Bing reports on, sourced from SwitchWize. Not a market measurement.
- The export is truncated. Bing returns roughly the top 29% of our citations by volume, not the full set. Relative ordering between queries is reliable; absolute totals are a floor, not a census.
- Citation share is Bing’s metric. Its exact definition is not fully public, so we report it as given rather than recomputing it.
- Share is not comparable across query sizes. It falls systematically as a query gets larger, because it is a proportion of every source cited and a bigger query draws on a bigger pool. Comparing a long-tail query’s share to a high-volume query’s share is a mistake we made ourselves before measuring it.
- Intent and topic labels are AI-generated by Bing. Useful signal, not a human-coded taxonomy.
- The window is short. Roughly three months, and our own series contains an unexplained multi-day dip that no change on our side accounts for.
- Citations are not visits. A citation means a page was used as source material. Whether anyone clicked is a separate measurement.
Questions
- What is a grounding query?
- When someone asks an AI assistant a question, the assistant issues its own search queries to find source material before writing an answer. Those are grounding queries. They are generated by the assistant, not typed by a person, so they are often longer and more literal than a human search, and they do not appear in conventional keyword tools.
- What does citation share measure?
- Bing reports, for each grounding query, what proportion of the sources used to compose answers came from a given site. A 20% citation share on a query means roughly one in five of the sources drawn on for that query was ours. It is Bing’s own metric and Bing has not published its exact formula, so we report it as Bing reports it rather than reinterpreting it.
- Does a citation mean somebody visited the site?
- No, and this is the most important distinction in the dataset. A citation means our page was used as source material for an AI answer. The reader may never click through. Citations and visits are separate measurements, and on our own data they diverge sharply: a large share of citations comes from pages that record no search impressions at all.
- Can another publisher reproduce this?
- Yes, for their own site. Any site owner with a verified Bing Webmaster Tools property can open AI Performance, view Grounding Queries, and download the same report for their own domain. Nobody can pull this report for a site they do not control, which is why cross-site comparison is not possible.
- Is this a measure of the whole AI search market?
- No. It covers Microsoft Copilot and the partners Bing reports on, sourced from one site. It says nothing about ChatGPT, Claude, Gemini, or Perplexity, and nothing about publishers other than us. Any characterisation of it as a market-wide measurement is wrong.
For journalists
We are happy to walk through the underlying export, explain how a specific figure was derived, or comment on the record. The per-query table is not published here, but we will discuss specific queries on request. Contact research@switchwize.com.
See also how we research everything else and our press page.