Microsoft Copilot

One index underneath several products

Microsoft Copilot is Microsoft's consumer AI assistant, and Copilot Search is the generative answer layer inside Bing. Both are grounded on the Bing index: Bing performs the crawl and the retrieval, and a language model writes an answer over the retrieved passages with inline links back to the sources. The name has moved twice. It launched as Bing Chat on 7 February 2023, became Copilot on 15 November 2023, and the brand now covers a consumer assistant, Copilot inside Windows and Microsoft 365, and Copilot Studio for building enterprise agents.

For a publisher, only the web-grounded consumer surfaces matter, and Copilot Search in Bing — launched worldwide except in Russia and China on 4 April 2025 — is the clearest of them. Microsoft describes it as blending traditional and generative search, citing its sources prominently, inline-linking whole sentences, and placing cited sources and web results at the top and bottom of the page.

Microsoft states the Bing grounding path most explicitly in its Copilot Studio documentation, where a generative answers node passes an optimized query to Bing Custom Search and the recommended publisher action is to use robots.txt to tell bingbot what it may access. The consumer assistant is not documented in that detail, but no separate consumer Copilot crawler is documented anywhere, which is a finding in its own right.

Why the reporting matters more than the traffic

Copilot is a small referral source and the honest way to introduce it is to say so. On SE Ranking's panel of 101,574 websites, covering January 2025 to April 2026 and published 18 June 2026, Copilot accounted for 3.51% of AI referral traffic, fourth behind ChatGPT, Gemini and Perplexity. On Similarweb's usage measure — traffic to generative AI websites rather than from them — it was 2.0% in May 2026, sixth. Both are small, and the two methods do not even agree on the ordering.

Dismissing Copilot on those numbers is the mistake. Microsoft is the only operator in this field that reports citations back to publishers, which means it is the only place where a publisher can observe generative citation behavior directly rather than infer it from clicks that may never come. The asymmetry is the story: Microsoft measures best what matters least by volume, and everybody else measures worst what matters most.

That has a practical consequence beyond Copilot. If you want to understand what grounding queries look like, how citation counts move week to week, or how a page's citation behavior differs from its ranking behavior, Bing Webmaster Tools is the only publisher-side instrument that exists. Google's AI Overviews and AI Mode report nothing separable, OpenAI publishes no impressions, and Perplexity publishes nothing at all.

What the AI Performance report shows, and what it cannot

The AI Performance report entered public preview on 10 February 2026. Microsoft describes it as showing how publisher content appears across Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations, with four things reported: Total Citations, Average Cited Pages, Grounding Queries, and page-level citation activity over time. On 16 June 2026 Microsoft added Intents, Topics, Compare, and Citation Share, defined as the percentage of citations attributed to your site out of all citations shown across all sites for a given grounding query.

Read Microsoft's caveats as carefully as its metrics, because they are unusually candid. Grounding-query data "represents a sample of overall citation activity." Citation counts reflect frequency, "not page importance, ranking, or placement." Citation patterns shift with user behavior, evolving models, freshness signals, partner refresh cycles and changes across the web. Topic labels "may still be broad." The features are described as early preview innovations.

Two misreadings are already in circulation. The first treats the report as cross-platform AI visibility: it is not, it covers Microsoft's own ecosystem plus unnamed partners, and it cannot substitute for measurement of ChatGPT, Gemini or Perplexity. The second treats Citation Share as share of voice or as traffic share: it is neither. It is an observational count for one grounding query, it exposes no competitor data, and Microsoft's own framing says so.

There is also a question the documentation does not answer. Both announcements use the phrase "select partner integrations" and neither names the partners. Whether a third-party assistant licensing Bing's index shows up inside your citation totals is therefore unknowable from what Microsoft has published, and it is the first thing a careful publisher would want to know.

NOCACHE and NOARCHIVE, the controls nobody else ships

Since 22 September 2023 Microsoft has offered page-level control over generative use, through two meta directives most robots.txt-focused advice never mentions. Microsoft's wording: content with the NOCACHE tag "may be included in Bing Chat answers. We will only display URL/Snippet/Title in the answer." Content tagged NOARCHIVE "will not be included in Bing Chat answers, not be linked to in the answers." The original announcement is still the primary source.

The granularity is what makes this different. A publisher can stay indexed for classic Bing results while limiting how the generative layer uses a given page, and can make that decision page by page rather than site-wide. OpenAI, Perplexity and Anthropic offer nothing comparable; their controls are crawler-level, all or nothing. Reporting in February 2026 on Bing's guidelines update described NOARCHIVE as preventing content from being used in Copilot responses and grounding results, so the 2023 mechanism is understood to cover the current products.

Anyone writing about AI content controls who lists robots.txt directives and stops has left out the only page-level generative controls any major operator has actually shipped, three years after they shipped.

There is no Copilot crawler to block

Publishers regularly go looking for a Copilot user agent to allow or disallow, on the reasonable assumption that Microsoft must have one because everybody else does. OpenAI has OAI-SearchBot, Perplexity has PerplexityBot, Anthropic has Claude-SearchBot. As of 29 August 2026 no Microsoft documentation of a consumer-Copilot user agent could be found: no agent name, no IP range file, no Copilot-specific robots.txt token. Everything routes through bingbot.

Bot directories do list assorted Copilot agent strings. The ones traceable at all lead to unattributed commercial directories rather than to any Microsoft page, and they should be ranked below every other class of source. Treat them as unverified until Microsoft publishes a user-agent list, or until server logs show a Copilot-named agent arriving from a Microsoft-owned address range.

The lever, then, is bingbot plus the page-level meta directives. Microsoft's own statement in the AI Performance announcement is that "Bing respects all content owner preferences expressed through robots.txt and other supported control mechanisms" — one sentence covering both, and worth noting that it contains no carve-out for user-initiated fetches of the kind OpenAI and Perplexity both publish.

Bing named GEO in its guidelines, and could not be quoted directly

On 27 February 2026 Bing added generative engine optimization — the practice of shaping content to be eligible for grounding and reference in AI answers — to its Webmaster Guidelines, and rewrote its language on machine-generated content. The old text defined machine-generated content mechanically; the new text targets large-scale content generated without oversight, quality control or editorial review, which is a judgment about process rather than about tooling. Bing also stated that GEO does not guarantee citations, just as SEO does not guarantee rankings.

An awkward detail belongs with that paragraph. Bing's Webmaster Guidelines page renders only with JavaScript and returns nothing to a plain fetch, verified 29 August 2026, so the primary text cannot be quoted from a fetched copy. Almost every account of this change quotes Matt G. Southern's report in Search Engine Journal instead, this page included. That is a second-hand chain, and it should be attributed as one rather than passed off as a quotation from Bing.

Naming GEO is not the same as documenting selection. Microsoft still publishes no criteria for what gets grounded or cited, no scoring description, and no equivalent of a quality rater's guidance. The recognition is real; the operational guidance behind it does not exist.

The study nobody has run

Ahrefs ran 15,000 long-tail queries across Google, Bing, ChatGPT, Gemini, Copilot and Perplexity in August 2025 and compared what each engine cited with Google's results. Copilot's overlap with Google's top 10 was 8.2%, mid-pack, well below Perplexity's 28.6%; across all six engines only 12% of AI-cited links appeared in Google's top 10 and roughly 80% ranked nowhere in the top 100.

Notice what that does not measure. The baseline is Google, and Copilot is grounded on Bing, so the study says nothing about whether Copilot cites the pages Bing ranks. No study measuring Copilot citations against Bing rankings could be found on 29 August 2026. That is a conspicuous gap, because Microsoft now publishes grounding-query data that would make the comparison possible for any publisher with Bing Webmaster Tools access and the patience to build the dataset.

Until somebody runs it, the widely repeated advice to rank in Bing in order to be cited in Copilot is an inference, not a finding. It is a plausible inference — the grounding index is Bing, and Microsoft says so — but plausible and demonstrated are different things, and this is a subject in which the difference has been expensive.

Frequently asked questions