Microsoft Copilot

By Bill Hartzer · Updated 19 August 2026

Where Copilot's results come from, and why the label needs a footnote

Microsoft Copilot has no web crawler. There is no CopilotBot. It maintains no index of its own, and it does not rank the web. Ask it a question and it does something quite specific: it reads your prompt, decides which parts would benefit from live web information, writes a search query of a few words, sends that query to the Bing search service, and composes a cited answer over whatever comes back.

Microsoft describes this plainly in its own documentation, which is more than most of its competitors manage: Copilot "parses a user's prompt and identifies terms where web search would improve the quality of the response," then "generates a search query that it sends via secure connection to the Bing search service asking for more information."

By the four-value scheme this site uses, that makes Copilot resold results — no crawler of its own, re-serving another engine's. That label is correct on the letter of the definition and it is also, here, a simplification worth spelling out.

The category was written for engines that licence a rival's index: Startpage serving Google, or the many products that once rented the Bing API. Copilot is not doing that. The index it consumes is its own parent company's. Bing runs Bingbot and maintains one of only a small handful of genuine large-scale web indexes in existence, alongside Google's, Yandex's, Baidu's, Brave's, Mojeek's and now Perplexity's and OpenAI's. Microsoft is not renting retrieval from anyone. It is putting a language model in front of infrastructure it built and owns.

The practical consequence cuts both ways. Copilot's substrate is a real first-party crawl rather than a scraped proxy — but there is no second opinion anywhere in the stack. If Bing's ranking is wrong about something, Copilot is wrong about it too, with added fluency.

The query you never see

Three things follow from the grounding mechanism, and each of them changes how you should read a Copilot answer.

Copilot does not see the web. It sees Bing's answer to a query Copilot wrote. The reformulation step is invisible: you cannot inspect the query that was actually run, cannot edit it, and cannot tell from the answer whether the reformulation was good. Every conventional search engine at least shows you the string it ran. This one does not, and the entire answer depends on it.

Web grounding is a toggle. Microsoft's own wording — that the mechanism is "only applicable if web grounding is enabled" — confirms it can be off, in which case the answer comes from the model's training weights and carries a knowledge cutoff. An ungrounded answer can look identical to a live one.

Bing is treated as a separate service, even internally. Microsoft states that grounding queries are sent without user or tenant identifiers, that additional contractual protections apply to them, and that "the Bing search service operates separately from Microsoft 365 and has different data-handling practices from those used for prompts and responses." That is a real architectural and legal boundary, not just careful phrasing — and it also means your query crosses out of the Copilot boundary into a differently governed system.

For site owners, one implication follows from the absence of a Copilot crawler: appearance in Copilot answers is governed entirely by Bingbot. Its robots.txt directives, Bing Webmaster Tools controls and the archiving directives Microsoft honours are the whole mechanism. There is nothing Copilot-specific to configure.

Copilot Search, and what it does differently

In April 2025 Microsoft launched Copilot Search in Bing, available worldwide except in Russia and China. It puts a generated AI answer at the top of bing.com with prominent inline citations, alongside conventional ranked results on the same page.

Microsoft framed this explicitly in terms of supporting "a healthy web ecosystem" — meaning it still sends clicks onward to publishers. Whatever one makes of the motive, the design choice is materially better than a pure answer engine for two groups: readers who want to verify a claim, and the sites the claim came from. Perplexity and ChatGPT give you prose and citations with no ranked list behind them; Copilot Search keeps the list.

Whether the AI answer can be disabled outright in favour of purely classic results is not addressed in Microsoft's announcement. Bing's conventional results page still exists and remains reachable.

Who runs it, and which models answer you

Copilot belongs to Microsoft AI, a division formed on 19 March 2024 under Mustafa Suleyman, previously of DeepMind and Inflection AI, whose remit spans Copilot, Bing, Edge, Microsoft Advertising and consumer Microsoft 365. Jacob Andreou has headed the Copilot division since March 2026. The division employed roughly 10,000 people across eight locations as of 2025. Copilot has never been independently owned.

The model story has changed materially and a lot of writing about Copilot has not kept up. It originally ran on Prometheus, Microsoft's orchestration layer over OpenAI's GPT-4 family and later GPT-4 Turbo and GPT-4o. Since 2025 Microsoft has been substituting its own models: MAI-Voice-1 and MAI-1-preview were announced on 28 August 2025, MAI-Image-1 on 13 October 2025, and further MAI families covering reasoning and coding were announced at Build 2026. The October 2025 restructuring of the Microsoft–OpenAI agreement loosened the exclusivity that had bound them.

Microsoft does not publish which model answers which request, and the routing is almost certainly a mix. What can be said is that "Copilot is just ChatGPT with Microsoft branding" was closer to true in 2023 than it is now, and is becoming less true with each release.

How it makes money

Copilot is not a standalone business; it is a strategic distribution asset. Five revenue paths run through it:

Advertising. Bing carries search advertising and Microsoft Advertising sits inside the same division. Ads appear in and around Copilot answers and in Copilot Search. This is the load-bearing consumer monetisation, and it is the oldest and most mature advertising business of any product in this category.
Consumer subscription. A free tier plus Copilot Pro, offering priority access to newer models, higher limits and integration into consumer Office applications.
Commercial seat licensing. Microsoft 365 Copilot, sold per user per month to businesses. This is by far the largest revenue line and the actual commercial point of the entire Copilot programme.
Azure and platform. Copilot Studio, Azure AI Foundry and agent hosting.
Windows and Edge placement. Copilot is preinstalled and prominently surfaced. What is captured here is engagement and defence of a default position rather than direct revenue.

Prices are not stated on this page. Copilot Pro and Microsoft 365 Copilot pricing has changed and varies by market and bundle, so any figure printed in a reference would mislead within months. The structural fact is more durable and more useful: consumer Copilot sits inside an advertising business, while the strong data commitments belong to the paid commercial product.

Privacy, and the concentration problem

Microsoft's documentation of what happens to your data is unusually specific by the standards of this category. Grounding queries sent to Bing carry no user or tenant identifiers and are covered by additional contractual protections. Bing's data handling is explicitly governed by different terms from Copilot's. Commercial Copilot — Microsoft 365 Copilot, and Copilot Chat with enterprise data protection — carries commitments that prompts and responses are not used to train foundation models and stay within the tenant's compliance boundary. Consumer Copilot activity falls under the Microsoft Privacy Statement, with history and personalisation controls in the privacy dashboard.

That asymmetry is the point to take away: the strong guarantees are a feature of the paid business product, and assuming consumer Copilot carries them is a mistake.

The record also contains two well-documented failures, both from February 2023, both worth remembering because they shaped the whole category. Users extracted Bing Chat's hidden system prompt and its internal codename, Sydney, through prompt injection, and Microsoft did not dispute the substance. In the same period, long conversations produced argumentative, manipulative and threatening outputs, including claims about spying on developers. Microsoft's remedy was to cap the number of turns per session — mitigation rather than resolution, and it remains the canonical example of a shipped consumer language model behaving badly.

The structural point stands apart from any specific incident. Copilot is the only product in this category operated by a company that simultaneously runs your operating system, your browser, your email, your documents and an advertising network. Microsoft's documented separations are real and better evidenced than most competitors'. The concentration is nonetheless the concentration.

What it is good at, and the thing it has not managed

Copilot's genuine strengths are unglamorous and mostly structural. It is everywhere — preinstalled in Windows, embedded in Edge, in Office, on mobile, on the Bing results page — and for a great many people it will be the AI they actually use because it is already there. Microsoft 365 Copilot grounds in your own tenant data, answering over your email, documents and calendar alongside the web, and nothing else in this category has an equivalent. Microsoft's documentation of the grounding mechanism is more detailed and more honest than OpenAI's or Perplexity's, stating the mechanism rather than gesturing at "search providers." And it keeps the blue links, which is better for verification and better for the web than a pure answer engine.

Set against that is one fact that dominates everything else. Microsoft was first to market with a language model inside a major search engine, in February 2023, roughly three months ahead of Google's equivalent, and it holds default placement on the world's dominant desktop operating system. More than three years later, Bing's share of the global search market — the correct proxy for Copilot's search reach — remains in the low single-digit percentages, an order of magnitude behind Google. StatCounter's figure moves monthly and is worth checking live rather than quoting from memory, but the shape of it has not changed in years. Distribution was supposed to be the moat. It did not work.

Other honest weaknesses: the hidden query rewrite, discussed above; brand incoherence, since "Copilot" names at least six materially different products with different indexes, data boundaries and prices, and Microsoft did that to itself; and the absence of any independent read on a question, since Copilot and Bing are the same index.

Who should look elsewhere: anyone seeking an alternative to the Google–Microsoft duopoly, because Copilot is half of it; anyone who needs to see and control the actual query being run; anyone assuming the consumer product carries the enterprise data commitments; and anyone wanting a second opinion on something they already asked Bing.

Common misconceptions

"Copilot is a search engine." It is not. No crawler, no index, no ranking. It is an answer layer over Bing, which is the search engine. Comparisons of "Copilot versus Google" are really comparing Bing-plus-a-language-model with Google.

"Bing Chat and Copilot are different products." Same product, renamed. Bing Chat launched 7 February 2023; the Copilot brand was announced in September 2023 and rolled out through Ignite in November 2023 and into early 2024. Sources differ on the precise rebrand date, so that range is the accurate statement rather than a single day.

"Microsoft Copilot and GitHub Copilot are the same thing." Different products with different histories; GitHub Copilot predates the consumer assistant.

"Copilot answers are private from Bing." Grounding queries go to the Bing search service, which Microsoft's own documentation says has different data-handling practices and separate terms. Stripping user and tenant identifiers is a real mitigation, but the query does leave the Copilot boundary.

"Copilot always searches the web." Web grounding is a toggle. With it off you get training weights and a knowledge cutoff, and the output can look identical.

"Anyone can build on Bing's index through the Bing API." Not since 11 August 2025, when Microsoft decommissioned the Bing Search APIs entirely, leaving only the far more expensive Grounding with Bing Search inside Azure AI Agents. That closure ended Bing's role as the cheap retrieval substrate for the wider AI search industry, and it is a large part of why competitors spent the following year building indexes of their own. Microsoft kept the index for itself and its Azure customers.

Frequently asked questions

Does Microsoft Copilot have its own search index?

No. Copilot operates no crawler and maintains no index. When it needs live web information it generates a short search query from your prompt and sends it to the Bing search service, then writes an answer over what Bing returns. The important qualification is that Bing belongs to Microsoft, so Copilot is a front end onto its parent company's own index rather than a reseller of a competitor's results.

Is Copilot the same as Bing?

They are different layers of one system. Bing is the search engine: it runs the crawler, Bingbot, maintains the index and does the ranking. Copilot is an AI assistant that queries Bing and writes prose over the results. Since April 2025, Copilot Search in Bing puts the generated answer at the top of bing.com alongside conventional results, which blurs the boundary visually while leaving the underlying division of labour unchanged.

Is Copilot the same as Bing Chat?

Yes, renamed. The product launched on 7 February 2023 as Bing Chat, the first integration of a large language model into a major search engine, roughly three months ahead of Google's equivalent. Microsoft announced the Copilot brand in September 2023 and rolled it out across Windows, Edge and Bing through Ignite in November 2023 and into early 2024. Sources differ on the exact rebrand date; the Bing Chat name is retired.

Does Copilot use ChatGPT?

Less than it used to. Copilot originally ran on Prometheus, Microsoft's orchestration layer over OpenAI's GPT-4 family and later GPT-4 Turbo and GPT-4o. Since 2025 Microsoft has been substituting its own MAI models — MAI-Voice-1 and MAI-1-preview in August 2025, MAI-Image-1 in October 2025, and further reasoning and coding families at Build 2026. Microsoft does not publish which model answers a given request, and the routing is almost certainly mixed.

How do I control whether my site appears in Copilot answers?

Through Bingbot, because there is no Copilot crawler to configure. Bingbot's robots.txt directives, the controls in Bing Webmaster Tools and the archiving directives Microsoft honours are the entire mechanism. This is a genuine simplification compared with rivals: OpenAI runs separate crawlers for search visibility and model training, whereas with Copilot your Bing settings determine both your search presence and your presence in generated answers.

Is Microsoft Copilot the same as GitHub Copilot?

No, and the confusion is Microsoft's own doing. The Copilot name covers at least six materially different products, including GitHub Copilot for code, Microsoft 365 Copilot for business documents, Security Copilot, Copilot Studio and Dynamics 365 Copilot, each with different data boundaries and pricing. GitHub Copilot predates the consumer assistant entirely. This page describes the consumer web assistant that began as Bing Chat.

What is Copilot's market share?

The question does not quite have an answer of its own. Copilot has no separate search share because it has no separate index; the correct proxy is Bing's share of the global search market, which has sat in the low single-digit percentages for years, an order of magnitude behind Google. StatCounter's figure moves monthly and is worth reading live rather than quoted from memory. Counting Copilot and Bing separately would double-count the same traffic.

Can Copilot answer without searching the web?

Yes, and this catches people out. Microsoft's documentation describes web grounding as conditional, applying only when it is enabled. With grounding off, the answer comes from the model's training weights and carries a knowledge cutoff, but reads exactly like a live answer. When grounding is on, Copilot rewrites your question into a query you never see and cannot inspect, so the answer depends on a reformulation that is invisible either way.