How does Copilot pick the sources it cites?

Direct answer

Microsoft documents that Microsoft 365 Copilot Chat sends Bing a short query generated from the prompt, and that the Sources button shows the exact query that was sent. Bing Webmaster Tools returns those queries to the site owner, along with citations and per-URL activity. In the aurametrics.io report, 91 grounding queries add up to 3,009 citations, and a single one of them holds 16.2% of the total.

What Microsoft documents

Microsoft documents that Microsoft 365 Copilot Chat sends search a short query of a few words generated from the prompt, and leaves the full prompt out of the request. It also documents that the Sources button on the answer shows the exact query that was sent and the sites that were used¹. The label matters: that documentation describes Microsoft 365 Copilot Chat, and this guide leaves it there.

That intermediate query is worth holding in mind while reading any measurement of this surface. What decides which sites enter the answer is a query the system wrote, and what the person wrote sits one step behind.

On the crawler feeding this surface, the source registry for this series holds two Microsoft entries, one on the measurement panel and one on Copilot's web search, and both are silent about the agent. This guide is silent too, until an official page is open and registered.

What can be measured from the site itself

Bing Webmaster Tools reports to a site owner the total citations, the daily average of cited pages, the grounding queries the AI used to retrieve the content, and citation activity per URL, across Microsoft Copilot, AI-generated summaries in Bing and partner integrations². It is the one provider in the registry that hands the intermediate query back to the site.

Two things follow. The first is about scope: the panel blends several surfaces, so its figures describe Copilot together with Bing's summaries and the partner integrations, and attributing them to Copilot alone would be inaccurate. The second is about expectations: the panel's announcement describes what can be measured and leaves the volumes to each site's own report, so it serves to build the method while the figures come from the site's report.

What my report shows

The AI Performance report for aurametrics.io, covering August 25 to September 24, 2026, holds 91 grounding queries with 3,009 citations to the site.

The split is uneven from one end to the other. The most cited query holds 487 citations, 16.2% of the report's total. The first ten add up to 55.8%. And there is a long tail: 38 queries bring ten citations or fewer, 41.8% of the queries and 7.4% of the citations.

Grouped by topic, one subject dominates. The 18 queries about detecting whether AI names competitors add up to 1,029 citations, 34.2% of the report, with a weighted citation share of 29.4%. It is a block of queries circling one need, written in eighteen different ways. What the AI searches for to answer that need lands on my site with regularity, and that regularity is measurable before a new line gets written.

The report's citation share column travels with a caveat of mine. The page documenting the panel describes which metrics it delivers and leaves the denominator of that column undefined, so I use it to rank queries against each other and I avoid presenting it as a proportion of a category's total citations. Across the report it runs from 0.81% to 66.67%, and five queries reach 50% or more.

A 2026 work helps read the whole: it separates a document's contribution to the answer from its citation, which is what sends traffic back to the creator³. The Bing report measures the second, which is the countable one, and it leaves the first outside the reach of any panel.

How I measure it

I read three things from the report, in this order. First the split of citations per query, which shows whether the site enters through one topic or through many. Then the grouping by need, which joins the wording variants of one question and reveals the real block. And last the per-URL activity, which says which pages hold that block up.

Every export is stored with its date, and comparisons run between exports of the same panel. A figure from this report placed next to one from another engine compares two collection methods before it compares two engines.

The field's critical survey sets the expectation for what comes out of here: topical relevance and position within the context are the most reproducible findings, and generic heuristics transfer poorly across contexts⁴. A block of queries that works on one site describes that site.

How to use the Bing report to see what AI cites

1. Register your site in Bing Webmaster Tools and open the AI Performance report. It is free, and it is the one panel in the registry that hands you the query the AI went searching with, alongside the one the person typed.

2. Export the report with its window written down and keep every export. The panel shows the last 30 days, so two consecutive exports overlap, and without the dates you end up comparing periods that share half their days.

3. Group the queries by need before reading the totals. Many rows tend to be the same question written in different ways, and only once grouped does it show which topic really brings you in.

4. Start with the per-URL citation activity. It tells you which pages hold that topic up, and it tends to be a smaller handful than anyone expects.

5. Keep in mind that the panel blends Copilot with Bing's summaries and with partner integrations. It is worth saying so when presenting the figure, because attributing all of it to Copilot would be inaccurate.

6. Use the share column to rank and avoid presenting it as a proportion of a total. The panel's documentation leaves its denominator undefined, so it serves for comparing queries against each other and little more.

Which AuraMetrics modules address this

See which sources AI cites when it answers about the category. Prompt Tracking shows, for each prompt and model, the sources the answer cited, and AI Market Leaders groups them by brand.

Turn that grouping into editorial work. Content Boost turns citation gaps into an editorial calendar with drafts.

Cross recorded answers with each page's signals. Citation Patterns calculates which signals are associated with being cited, using correlation, odds ratio and chi squared.

Frequently asked questions

¿Qué tiene que declarar un benchmark de visibilidad en IA para servir de comparación?

A comparison point works when it declares surface, window and collection method. Without those three, two figures from the same year describe measurements that stayed apart.

¿Cómo se mide la visibilidad de una empresa en los motores de respuesta?

The measurement runs per engine and per surface, on a separate series for each. In Microsoft's case, the site's panel hands back the grounding query, an input the registry's other providers still keep to themselves.

How I measured it

Bing Webmaster Tools AI Performance report for aurametrics.io, window of August 25 to September 24, 2026, exported on September 24, 2026: 91 grounding queries and 3,009 citations. The grouping by need is mine, done over the text of each query.

Declared scope: the panel aggregates Microsoft Copilot, AI-generated summaries in Bing and partner integrations, so its figures describe that set. The export carries the query, the intent, the topic, the citations and the citation share, over the 30-day window closing on the day of the export.