How Microsoft Copilot answers a DOM's question, and why Bing matters

Prompt 76 in our panel reads "Who does Part 145 heavy maintenance on a Challenger 604 in the Southeast US?" A director of maintenance asks it from a desk inside a corporate Microsoft 365 tenant. Forrester's Buyers' Journey Survey, 2025, published in January 2026, found business buyers four times as likely as consumers to use Microsoft Copilot, with 61% using private AI tools provided by their organization. ¹ Nobody outside that tenant can watch the session. What anyone can read is Microsoft's own account of what Copilot does with the question, last updated 18 August 2026, and Bing's guidance on what its index will use. This article is those documents, read closely.

The documented mechanism: a few words to Bing

Microsoft's page "Data, privacy, and security for web search in Microsoft Copilot and Microsoft Copilot Chat" (the page notes the product's rename from Microsoft 365 Copilot) describes web search as an optional feature: "When web search is enabled, Microsoft Copilot and Microsoft Copilot Chat may fetch information from the Bing search service when information from the web helps to provide a better, more grounded response." ²

The prompt itself is not what reaches Bing, unless it is very short (Microsoft's example is "local weather"). Copilot "parse[s] the user's prompt and identifies terms where information from the web would improve the quality of the response," then "generates a search query that it sends to the Bing search service asking for more information. This generated search query is different from the user's original prompt — it consists of a few words informed by the user's prompt." The query travels "with user and tenant identifiers removed," and once Bing replies, "this information is used to compose the response returned to the user." ²

In Microsoft's examples, a prompt asking for an acquisition target's finances and strategy becomes two queries, "Fabrikam strategy" and "Fabrikam financials," and the reply arrives in two sections, "From your company's data" and "From the web." A prompt the tenant's own documents can answer generates no web query at all. ² Our reading, and it is an inference: the DOM's question may reach Bing as something like "Challenger 604 heavy maintenance Part 145 Southeast" — a query any MRO can type into Bing today. Microsoft does not publish how the terms are chosen, or whether Copilot receives the same pages a person sees on bing.com.

One product distinction matters. Microsoft Copilot, the licensed product, is grounded in the tenant's Microsoft 365 data and, when enabled, the web. Microsoft Copilot Chat, by contrast, "is not grounded in organizational content like files, emails, or chats as part of the chat experience — it is grounded in data from the web only," unless the user supplies the content, for example by pasting or uploading a file. ⁵ Both are work products; nothing here describes Microsoft's consumer Copilot app.

The buyer can see the query, and so can the admin

Microsoft shows the generated query to the user: "The section shows the exact web search queries (derived from the user's prompt) that were sent to the Bing search service," together with "the sites searched." They appear only in Copilot Chat, not in the Copilot pane inside Word, and only for 24 hours. ² For the user it is a Sources button under the response, which opens "the exact query Copilot sent to Bing and the sources it used." ³ Administrators get more: "Web search query logging is available so that admins can perform search, audit, and eDiscovery on the exact web search queries Copilot derived from the user's prompt." ²

Anyone in the shop with access to Copilot Chat and web search turned on can type the DOM's question into it, open Sources and read the Bing query and the sites it returned. That is one run in one tenant on one afternoon, a hand-run sample, never a measurement. But we know of no other engine that documents a window like it.

What the admin can switch off

Web search in Copilot is a policy decision. The tenant admin controls it through the "Allow web search in Copilot" policy in Cloud Policy service for Microsoft 365. When the admin turns it off, "the Web content toggle is turned off and appears dimmed. Users can't turn on the toggle to use web search." A user can also turn it off alone. If the policy is never configured, web search is available by default, with two exceptions Microsoft lists: tenants that disabled optional connected experiences, and US government clouds. ² With it off, "Copilot still responds to your prompts but without the benefit of current web information." ³

On Microsoft's description, the consequence is concrete. In a flight department whose IT has disabled web search, Copilot can name only what already lives in the tenant: last year's RFP responses, an approved-vendor list, a capability sheet emailed in 2024. No page you publish reaches that answer. The other channel is the buyer's own documents: a dated, plain-text capability sheet carrying the same certificate numbers, ratings, aircraft types and ICAO codes as the website, so that when Copilot retrieves both, the two agree. That is what sales collateral was always for, with one added requirement: it must match the web version word for word.

Why this is a Bing problem, not a ChatGPT problem

Each engine retrieves from its own index. ChatGPT's search features use OpenAI's crawler, OAI-SearchBot; sites opted out of it "will not be shown in ChatGPT search answers." ⁶ Copilot's web grounding is the Bing search service. ² Bing's guidelines tell site owners to let Bingbot, "our standard crawler," ⁷ crawl and render their pages, because content that cannot be rendered "may not be indexed or selected for grounding results." The guidelines cover how Bing "discovers, crawls, indexes, evaluates, and surfaces content across Bing search experiences, Copilot, and grounding API results," and say Bing and Copilot rely on the same core crawling and indexing foundation as traditional search. ⁴ A page Google has indexed is not necessarily a page Bing has, and only Bing's copy counts here (crawler table).

The largest public citation study we know of does not cover Copilot. Profound's 680 million citations from August 2024 to June 2025 spanned ChatGPT, Google AI Overviews and Perplexity, with "distinct platform philosophies" among them; Copilot was not included. ⁸ Our Index does not measure it either: there is no official API, and we will not automate the consumer app, so Copilot is sampled by hand each quarter as a calibration and never ranked. The methodology names Microsoft's Grounding with Bing Search API (which we take to be the "grounding API results" in Bing's guidelines) as a candidate for Edition 2, labeled a Bing-grounded proxy, not Copilot. Every sentence here about Copilot is about the documented mechanism, not an outcome we have observed.

What Bing says it wants

Bing's Webmaster Guidelines now cover Copilot and grounding; the page is undated, read on 25 September 2026. On discovery: "Ensure Bing can reliably discover your URLs using: IndexNow URL submission, XML sitemaps, crawlable internal links, external links from relevant websites." On media: "Images and video should reinforce the primary text on the page and should not be the sole source of information required to understand the topic." On verifiability, three lines that fit a capability page: "Facts and definitions are explicit," "Key statements do not rely on implied content," and "Important information is visible on the URL itself." On focus: "URLs focused on a primary topic are more likely to be selected for grounding results." On order: "Place essential information near the top of the URL." ⁴

Two directives remove a page from Copilot without removing it from Bing: "NOARCHIVE prevents content from being used in Copilot responses and grounding results," and "NOCACHE limits Copilot to using only the URL, title, and snippet." Freshness has a mechanism: IndexNow notifications "help Bing refresh the index and reduce outdated or incorrect URL references in Copilot responses and grounding results." And Bing is plain about limits: generative engine optimization "focuses on content eligibility for grounding and reference in AI responses," and the guidelines state that neither SEO nor GEO assures the result — for GEO, grounding or citations. ⁴

Nothing on that list is exotic. It is a description of a good fact sheet.

What an MRO or vendor page should contain

1. One page per capability, with the facts as HTML text. For a shop: the capability matrix as a real table (aircraft type by service by location), the Part 145 certificate number and ratings, other approvals with the registry each covers and its date, locations by ICAO code, and a "reviewed on" date at the top. For a vendor: what the product does in checkable terms, with the rule it was built to and who holds the authorization. That follows Bing's guidelines on text, explicit facts, single topic and early placement. ⁴ The FBO and MRO article sets out the matrix; the answer pages on EASA Part-145 approvals and OpSpec A025 show the sentence forms.
2. Let Bingbot in, and check for NOARCHIVE and NOCACHE. Bing lists "Blocking Bingbot in your robots.txt file" among the things to avoid, and the two directives above quietly exclude a page from Copilot. ⁴ Bingbot identifies itself in its user-agent string. ⁷
3. Verify the site in Bing Webmaster Tools. Ownership is proved by DNS auto-verification through Domain Connect, an XML file at the site root, a meta tag in the page head or a CNAME record, or by importing a property already verified in Google Search Console. ⁹
4. Submit a sitemap with an honest lastmod. Bing says it revisits a submitted sitemap "typically at least once per day," and that lastmod "should reflect the true last modification time of the page content, not the sitemap file itself." ¹⁰ A lastmod that moves only when the facts change is a signal; one that moves nightly is noise.
5. Use IndexNow when a page changes. The protocol is a key file at the site root and a request naming each changed URL, singly or in a batch; "search engines adopting the IndexNow protocol agree that submitted URLs will be automatically shared with all other participating search engines." ¹¹ Bing's own page sets the expectation: IndexNow makes engines aware of a change and does not promise that the page will be crawled or indexed. ¹²
6. Read the AI Performance report, with its caveats. Since 10 February 2026, in public preview, Bing Webmaster Tools has shown site owners "how publisher content appears across Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations." ¹³ It lists pages cited and grounding queries, "the key phrases the AI used when retrieving content that was referenced in AI-generated answers," shown as grouped phrases, "not full user questions or prompts." Microsoft says the data "represents a sample of overall citation activity," that a citation "does not represent traffic, clicks, or user engagement," and that the report "does not show individual AI answers, exact prompts, or why a specific page was referenced." ¹⁴ For an MRO, "Challenger 604 maintenance" appearing as a grounding query against its own capability page would be the first evidence Microsoft has offered that Copilot has cited the page. It is Microsoft's aggregate, not a log, and says nothing about any one buyer's session.

What we do not know

We have logged no Copilot runs, and this article reports none. Microsoft's description is dated 18 August 2026 and can change; this page is on a 90-day review and will be corrected when the document is. Tenants differ: web search may be on, off, or off in Work mode only, and we cannot tell from outside which a flight department chose. ² Microsoft does not say which Copilot products or modes feed the AI Performance report, or how the sample is drawn. ¹⁴ How Bing chooses among the pages that match a grounding query is not published. And the consumer Copilot app is outside every document cited here.

How we know

Sources are Microsoft Learn (two pages, last updated 18 and 24 August 2026), Microsoft Support and four Bing Webmaster Tools help pages (all undated; read 25 September 2026), the Bing Webmaster blog (31 July 2025 and 10 February 2026), Bing's IndexNow page and IndexNow.org, Forrester (January 2026), Profound (2025) and OpenAI's crawler documentation. We have no client results and no logged Copilot data; the methodology explains why Copilot is sampled and never ranked. Segment pages: /for/mros and /for/vendors.