Insight · July 4, 2026
How Copilot
chooses sources.
Copilot does not read the whole web to answer you. It rephrases your question, pulls a short list of pages from Bing, and cites the few it leaned on. Everything about visibility follows from that.
01 · The shift
The engine now shows you the citations it once hid.
Microsoft spent 2026 turning Copilot citation behavior from a black box into a report. In February it shipped the AI Performance view inside Bing Webmaster Tools as a public preview. In June it added Intents, Topics, Citation Share, and Compare. For the first time a publisher can see which pages Copilot cites, which internal queries pulled them in, and how much of the citation space a site holds for a given query.
That reporting matters less than what it admits. Copilot is not ranking pages in a list and waiting for a click. It retrieves a handful of sources, reads them, writes one answer, and names the few it used. The report is Microsoft telling you, in its own words, that this is the surface you compete on now.
02 · The mechanism
Your question is not the query Copilot runs.
When you ask Copilot something, it does not search Bing for your exact words. It rephrases your question into one or more internal search phrases, built for retrieval rather than for a person to type. Microsoft calls these grounded queries. They are the real queries. The words you typed never touch the index directly.
That rephrasing sits on top of Prometheus, the technology Microsoft built to combine the Bing index with the reasoning of OpenAI GPT models. A component Microsoft named the Bing Orchestrator generates the internal queries, retrieves results from the Bing index, and hands them to the model. The model reasons over that retrieved data, which is what Microsoft means by grounding, then writes an answer with citations woven into the sentences.
The consequence is simple and strict. If a page is not in the Bing index, it is not in the candidate pool. If it is not in the candidate pool, it cannot be cited, no matter how good it is.
The one line to keep
“Copilot does not cite the best page on the web. It cites the best page it retrieved.”
03 · From a question to a citation
Five steps, every time you are cited.
01
Rephrase
Copilot turns your question into one or more grounded queries, internal search phrases built for retrieval, not the words the user typed.
02
Retrieve
Each grounded query runs against the Bing index. A page that is not indexed for that query is never in the running.
03
Read
The model reads the pages it pulled and reasons over them. This is grounding. The answer is anchored to what was retrieved, not to what the model remembers.
04
Write
It synthesizes one answer and weaves citations into the sentences, naming the few pages it leaned on most.
05
Report
Bing Webmaster Tools now shows the result back to you: the grounding queries, the cited pages, and your share of the citations.
04 · The new unit of visibility
Grounding queries are not your keyword list.
The old model gave you a keyword and a rank. The new model gives you a question you never see and a share of a citation you have to earn. In June 2026 Microsoft started classifying grounding queries by intent inside the report, into groups like informational, commercial, navigational, research, and local. It also added Citation Share, the percentage of all citations for a grounding query that point to your site.
Read together, these say something a rank number never could. You can be cited and still hold a thin share, because Copilot names several sources in one answer. You can cover an intent you never wrote a page for, because the model rephrased its way there. The work is no longer picking keywords. It is understanding the questions the model asks on your behalf and being the clearest answer to them.
05 · One grounding query, several named sources
There is no page one. There is a paragraph with three sources.
Picture a person asking Copilot which project management tool fits a small design studio. That is not the query the index sees. Copilot rephrases it into a retrieval phrase, runs it against Bing, and pulls a short list of pages that answer it directly. Then it writes one paragraph and cites a few of them.
Your page is competing for a slot in that paragraph, against every other page Bing indexed for the same grounded query. Being on the first screen of a results list is not the goal, because there is no list. Being one of the sources the model trusted enough to name is the goal. The report below is what Microsoft now shows for exactly that.
grounding query which project management tool for a small studio intent commercial cited pages several named sources in one answer citation share the percent of those citations that point to you
A grounding query maps to cited pages · the unit of AI search visibility
06 · What actually moves your odds
Be in the index. Answer cleanly. Be confirmed elsewhere.
Start with the index. If Bing does not have your page, nothing else matters. Confirm it is indexed and retrievable for the questions you care about. This is unglamorous, and it is the whole foundation.
Then write for retrieval. A page that answers one question cleanly, in plain language, near the top, is easier for the model to lift and cite than a page that buries the answer under a scroll of context. Clear headings, direct answers, and honest data give the model something to ground on and something to quote.
Corroboration is the quiet signal. The model favors a claim it can see confirmed in more than one place. A number that appears only on your own page is a risk. A number other indexed sources repeat is safe to cite. Freshness matters where the question is time sensitive, because the model reaches for the Bing index precisely when its own memory is stale.
None of this is a trick. It is the same thing that made a page worth citing before the model existed. The difference is that the model now does the citing, and it does it at the level of a single question.
07 · Where this does not apply
One engine, one index, an early report.
Enterprise Copilot grounded in a company own documents is a different system. It retrieves from an internal index, not the public web, so none of this reaches it. If your audience lives inside a tenant, public visibility work will not follow them there.
Reporting is also young. The AI Performance view is a preview, the numbers are aggregated, and Microsoft is still adding to it. Treat the citation share and the grounding queries as direction, not as a ledger. They tell you which questions you are answering and which you are missing. They do not promise a click.
And Copilot is one engine. The mechanics here are specific to the Microsoft stack, Prometheus, the Bing index, the Bing Orchestrator. The instinct carries to other answer engines, but the plumbing does not. Each one grounds differently.
Closing
You are not ranking anymore. You are being retrieved.
Open Bing Webmaster Tools and read your grounding queries. They are the questions Copilot already asks on your behalf. Answer the ones you are missing, clearly enough to be the source it names, and you are doing the only work that moves an AI citation.
Microsoft Bing blog · Introducing AI Performance in Bing Webmaster Tools, February 2026, and New AI Visibility Insights, June 2026 · Prometheus and the Bing Orchestrator per Microsoft Building the New Bing · reporting corroborated by Search Engine Journal and independent AI visibility analysts
Share this perspective
More insights
Adjacent perspectives.
Bttr. Field Brief
The brief Bttr. writes for senior buyers.
Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.