Blog · Engines

Microsoft Copilot and the Bing index: what it means for visibility

Published 31 August 2026 · Updated 8 October 2026 · 7 min read

Microsoft Copilot and the Bing index: what it means for visibility
Photograph: İsmail Enes Ayhan / unsplash

The short answer: Copilot builds its answers from Microsoft's grounding layer, which sits on the Bing index, so being in that index is the entry ticket and nothing more. The contest for a citation is decided by queries Copilot writes for itself, not by the question your buyer typed. Since February 2026 Bing Webmaster Tools has shown site owners those internal queries and the pages cited, and since June 2026 the share of the citation space a site holds.

That last part is the reason this engine deserves its own post. We read the published webmaster documentation of the engines we track on 8 October 2026 and found no equivalent first-party record of why an assistant fetched a page. That is what the documentation shows, not proof that nobody else has one.

The chain that ends in a citation

Picture the path a question takes. Somebody asks Copilot which payroll tool handles contractors in three countries. Copilot does not run that sentence as a search. It converts the intent into one or more retrieval queries, pulls candidate documents from the grounding system, selects passages, and writes an answer with a few links attached. Microsoft calls the retrieval step grounding, and in February 2026 it described grounding as the connective layer between models and current information, claiming that its own grounding serves nearly every major assistant on the market.

Four things have to go right for you, in order. Bing has to hold a current copy of the page. The internal query has to look like your page. A passage inside the page has to answer that query cleanly. The answer writer has to choose you over the other candidates it kept. Miss the first and the other three never happen, which is why access work comes before clever writing. The ordering follows from the shape of a retrieve-then-write pipeline rather than from anything specific to Microsoft, and how each AI engine picks the sources it cites sets out how far it carries to engines that document less.

The queries are not your prompts, and the gap is the finding

Bing Webmaster Tools exposes what it calls grounding queries: the phrases its systems used when retrieving content that ended up referenced in an answer. They are machine-written. A single user question can fan out into several of them, each narrower and flatter than anything a person would type.

Export a month of yours and put it next to your keyword list. In the client exports we have read since the report opened in public preview, a handful of domains and nothing like a census, a noticeable part of the grounding queries sat outside the keyword list, the prompt set and the content briefs. We have not counted them in a way worth quoting a rate from, so treat that as a pattern to check on your own export rather than a finding. The phrases that surprised people were specification-shaped and comparison-shaped, and often named a feature the site treats as a footnote.

Microsoft layered two more views on top of those phrases on 16 June 2026. Intents sorts the queries into categories such as informational, commercial, navigational, research and local. Topics clusters related queries into themes, on the argument that these systems reason over subject areas rather than isolated strings. Both are classifier output, and Microsoft says plainly that labels may be broad during preview. Treat them as a sketch of how the engine files you, not as a verdict.

Citation Share is a measurement, not a scoreboard

The June release also added Citation Share: the percentage of all citations shown for one grounding query that belong to your site. Microsoft defines it as the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query. We know of no other large engine that publishes a relative visibility number from its own logs rather than from sampled probing, though we can only go on what each engine documents in public.

Read the fine print before it goes into a board deck. Microsoft describes the metric as observational. It does not name the other domains, it is not traffic share, and it carries no quality judgement. Your share can fall while your citations rise, because the denominator is everybody else's work. It can also move because a model shipped, a freshness cycle ran, or demand shifted. A dip is a question to investigate, never a conclusion.

The grounding query data is also described as a sample of overall citation activity, not a census. Rates computed on a sample of unknown size are not rates. Pair the report with your own repeated prompt runs, and read how many runs per prompt before a number means anything before anyone quotes a percentage change at you.

The index is wholesale, and that is what most people miss

Here is the part that changes budgets. Microsoft now sells access to that grounding stack as a product. The Bing Search APIs were retired on 11 August 2025 and customers were pointed at Grounding with Bing Search inside Azure AI Agents. On 2 June 2026 Microsoft launched Web IQ, a set of grounding APIs it describes as a search engine for AI systems, built on the Bing global index and re-architected for agents.

Two consequences follow. The first is upside: a single crawl relationship can feed answers beyond Copilot's own surface, to the extent that an assistant or agent buys grounding instead of crawling the web itself. Microsoft's claim that its grounding powers nearly every major assistant is the vendor's own, and a publisher cannot verify it from outside. The second is the risk, and it is the familiar one. A bingbot block, a firewall rule that trips on datacentre ranges, or a client-side render that returns an empty document does not cost you one assistant. It removes you from whatever products are grounded on that index, and Microsoft does not list them, so you cannot enumerate the blast radius from outside. The Webmaster Tools report shows the loss as a decline with no cause attached. Microsoft states that Bing respects content owner preferences expressed through robots.txt and other supported controls, which means this is a setting somebody owns, not weather.

Being discoverable through someone else's index is still rented ground, and the arithmetic of that is the whole argument in rented visibility vs owned visibility in AI search. Rent it. Just know the lease terms.

A check you can run this week

Four steps, about ninety minutes, no tooling beyond a spreadsheet.

  1. Verify the domain in Bing Webmaster Tools and open the AI Performance report. Pull the

    2. last thirty days of grounding queries and cited URLs.

  2. Put your own buyer prompt list beside the grounding queries. Count the machine queries

    3. you had never written down. That count is your blind spot, expressed as a number.

  3. Compare the cited URL list with the pages you believe are your answer pages. When a

    4. pricing table or a support article outranks the page you wrote for this purpose, the engine is telling you where the clean passage lives.

  4. For the top five grounding queries, fetch each cited page the way a bot would, with

    5. JavaScript off, and confirm the answering sentence is in the raw HTML.

Google's Search Console documentation described no equivalent grounding-query report when we checked it on 8 October 2026, so the same exercise there needs inference rather than logs, which is the subject of how Google AI Overviews chooses its sources. If you want the longer version of this method with the evidence attached to each finding, the sample audit report shows the shape we use.

Honest limits

This post establishes what Microsoft documents and what we have read in client exports, and not much more. We do not know the sampling rate behind grounding queries, so absolute counts cannot be compared between two sites. We cannot tell which partner surfaces are inside the aggregate, because the documentation says Copilot, AI summaries in Bing and select partner integrations without listing them. Our claim that no other large engine publishes this kind of report rests on reading public documentation on 8 October 2026, which cannot rule out a report we have not found or one shipped since. The observation about grounding queries falling outside a client's keyword list comes from a small number of exports we read by eye, with no counting protocol behind it, so it is a prompt to check your own data and not a rate. And there is still no clean line from a citation to revenue, since these surfaces send little or no referrer.

FAQ

Does ranking well in Bing mean Copilot will cite me? No. Bing coverage is a precondition, not a cause. Citation is decided per grounding query, and a page that sits outside the top ten blue links can still carry the cleanest passage for a narrow machine query.

Is the AI Performance report free? Yes. It arrived in public preview on 10 February 2026 for verified sites in Bing Webmaster Tools, and the Intents, Topics, Citation Share and Compare views followed globally in preview on 16 June 2026.

Why did my citation count drop without any site change? Citation activity moves with model updates, freshness cycles, competing content and shifts in what people ask. Check crawl health and robots rules first, then compare periods rather than days, then look at whether the grounding queries themselves changed.

Should I block AI crawlers to protect my content? That removes you from the answers built on that index, and in Microsoft's case the same grounding reaches products well beyond Copilot. For a paywalled archive it can be the right trade. For a product site it rarely is.

Sources

Related reading

How each AI engine picks the sources it citesHow Google AI Overviews chooses its sourcesHow many runs per prompt before a number means anythingRented visibility vs owned visibility in AI searchWhy five engines give five different answers about youBrand mentions and citations in Gemini answers