What Sources Does Copilot Cite Most? A 2026 Q4 Data Study

Ask Microsoft Copilot which robot vacuum to buy and the answer will probably cite Amazon, Walmart and Best Buy. Ask it how to roll over a 401(k) and the citations shift to Forbes and NerdWallet. Ask it a question about your own software category and, unless you have looked, you have no idea whose pages it is quoting. That last question is the one Cognizo is built to answer: it tracks which domains and pages Copilot cites for the prompts your buyers actually type, every day, next to the other major answer engines.

This study pulls together what the public data says about Copilot citation sources as of Q4 2026. We combined seven datasets published between February and October 2026, covering more than 3 million US Copilot queries, a peer-reviewed audit of 712 real-world questions, a finance-specific index of 1,664 Copilot responses, and first-party Bing Webmaster Tools data from two publishers. Where the studies disagree, we say so.

Key findings

• Retail owns the top of the list. In Ahrefs' September 2026 snapshot of more than 3 million US queries, Amazon (16.8%), Walmart (12.6%) and Wikipedia (7.6%) are the three most-cited domains. Retail and marketplace domains together hold about 50% of the citation share among Copilot's 50 most-cited sites.
• Reference sites are the second pillar. Wikipedia, Grokipedia, wikiHow, Britannica and three dictionaries add up to roughly 16.6% of top-50 share.
• Copilot cites more AI-generated pages than any other engine tested. A May 2026 audit accepted at AIES 2026 found that 27.8% of Copilot's cited sources were likely AI-generated, against 14.7% for Gemini, 9.4% for Perplexity and 7.3% for ChatGPT.
• Reddit is close to invisible. Copilot was the only engine in cloro's 2026 data that cited LinkedIn more often than Reddit (0.67% of answers vs 0.00%), and 81% of those LinkedIn citations pointed at long-form LinkedIn articles.
• Your industry changes the whole picture. In finance, Forbes appears in 20.73% of Copilot answers and only 36 domains overlap between the finance top 100 and Copilot's overall top 100.
• Citations pile onto very few pages and move fast. One publisher saw 69% of its 19,717 Copilot citations land on a single page, then watched daily citations fall 97% in two months.

How we built this Copilot citation study

No single public dataset answers "what does Copilot cite?" for every type of question. Large panels measure consumer queries well but say little about B2B. Webmaster data is precise but covers one site at a time. So we treated each source as a different lens and kept their numbers separate rather than averaging them into a figure none of them measured.

Two things to keep in mind while reading. First, the metrics differ: Ahrefs reports a share of citations, while cloro and LLM Authority Index report the share of answers that cite a domain at least once. A 1% citation share and a 1% answer share are not the same thing. Second, most of these studies come from companies that sell AI visibility tools, Cognizo included. That is why we link every number to its source.

The 20 most-cited websites in Microsoft Copilot

Here is the top of Ahrefs' September 2026 list. "Share" is each domain's citations as a percentage of all citations going to the top 50 domains, so it shows how Copilot splits its attention among its favorite sources.

Further down the list sit Mayo Clinic (1.1%), Cleveland Clinic (1.0%), Healthline (0.9%), Consumer Reports, LinkedIn, BBB, Trustpilot, WebMD and Britannica (0.8% each). LinkedIn is the only professional network in the top 50, at rank 37.

When we group the 50 domains by type, the shape is clear:

These groupings are ours, built from Ahrefs' published list. Shares are rounded to one decimal in the source, so the totals are approximate.

Finding 1: Copilot answers shopping questions with stores

Half of the top-50 share going to retailers says a lot about what people ask Copilot. Ahrefs notes that Microsoft is adding shopping and checkout inside Copilot chats, and the citation list looks like a product built for that. When the question is "best air fryer under $100," a product listing page with price, specs and reviews answers it in one hop.

For brands that sell physical products, this cuts two ways. Your own product pages compete with marketplace listings of the same product, and the marketplace usually wins on review volume. But it also means your Amazon, Walmart and Target listings are effectively part of your Copilot presence. Titles, bullet points and review counts on those listings are being read as evidence.

For B2B and service companies, the retail dominance mostly tells you what to ignore. Your buyers' prompts never touch Amazon, so the overall top 50 is the wrong benchmark. The vertical sections below are closer to your reality.

Finding 2: Encyclopedias and dictionaries are the default fallback

Wikipedia at 7.6% is expected. Grokipedia, xAI's AI-written encyclopedia, at 3.1% and rank 6 is less expected. Add wikiHow, Merriam-Webster, Collins, Cambridge and Britannica, and reference content takes about one sixth of the top-50 share.

The practical lesson is about definitions. If your category has a term that buyers look up ("what is a headless CMS"), Copilot will reach for a reference page first. A brand can still win that citation, but only with a page that defines the term in the opening lines as cleanly as an encyclopedia would, and then goes further than the encyclopedia can.

Finding 3: Copilot leans on AI-generated and low-authority pages

This is the finding that should make content teams pay attention. In the peer-reviewed audit by Mowafak Allaham and Nicholas Diakopoulos, cited pages were scraped from each engine's interface and scored with the Pangram detector, counting only pages rated "likely" or "highly likely" AI-generated:

The authors call their detector setup conservative, so the real shares may be higher. Copilot's figure is almost four times ChatGPT's.

The Ahrefs list points the same way. Three of the top ten domains are general explainer sites (scienceinsights.org, engineerfix.com, biologyinsights.com) rather than established publishers. Seven domains in the top 50 have an Ahrefs Domain Rating under 40, including one at DR 9 and one at DR 5, and together they hold about 6.5% of top-50 share. We have not tested whether those specific sites are AI-written, so treat the overlap as a pattern worth watching, not a verdict.

What it means: Copilot does not seem to reserve citations for famous domains. A focused page that answers the grounding query directly can get in, even from a small site. That is good news for challengers and a warning for category leaders, because a thin page from a competitor or a content farm can take the citation you assumed was yours. The only way to know is to watch the actual answers.

Finding 4: Reddit barely shows up, LinkedIn articles do

On most answer engines, Reddit is one of the most-cited domains. Copilot is the exception. Profound's analysis of billions of AI citations, published in November 2025, ranked Reddit 31st on Copilot, the lowest position among the six engines it compared, and noted that Copilot tends to cite Reddit next to business sources like Forbes and official forums.

cloro's 2026 data is even starker. In 55,479 Copilot answers to commercial prompts in the 30 days to August 7, 2026, Reddit appeared in 0.00% of answers. LinkedIn appeared in 0.67% and YouTube in 0.78%. Copilot was the only engine in the study that cited LinkedIn more than Reddit, and 81% of its LinkedIn citations went to long-form LinkedIn articles (the /pulse/ URLs), not feed posts. By the October window, LinkedIn's share had dropped to 0.36% of 65,038 answers, but articles still made up 212 of the 232 answers that cited LinkedIn at all.

Microsoft owns both Copilot and LinkedIn. cloro is careful to say its data cannot prove that ownership causes the pattern, and we agree. The actionable part is narrower: if your team writes thought leadership, a LinkedIn article has a citation path into Copilot that a LinkedIn post almost never does. Our own LinkedIn citation research and analysis of Reddit in AI answers cover how those two platforms behave across the other engines.

Finding 5: Each industry gets its own Copilot

The overall top 50 is dominated by shopping because shopping queries dominate the sample. Narrow the questions and a different set of sources takes over.

Finance. LLM Authority Index tracked 1,664 Copilot answers across 11 investing, retirement and mortgage categories from July to September 2026. The most-cited domains, by share of answers citing them at least once:

Only 36 domains appear in both the finance top 100 and Copilot's overall top 100. Specialists punch far above their general weight: StockBrokers.com shows up about 4.37 times as often in finance answers as it does across Copilot overall. The leader also changes by sub-topic. U.S. News leads annuities, NerdWallet leads mortgage refinancing, and StockBrokers.com leads online brokers.

Beauty. A study by Maverick Indonesia of 8,000 AI answers and 63,286 citations in August 2026, reported by Marketing-Interactive, found that news media supplied 53.1% of Copilot's citations on beauty prompts, while social, blog and community content supplied 1.8%. Across all five engines in that study, beauty brands controlled only 11.6% of the cited sources. The sample is Indonesian consumers, so the exact numbers will not transfer to other markets, but the direction matches the rest of the data: Copilot prefers editorial and reference sources over community chatter.

B2B software and services. Public panel data is thinnest here, which is exactly why we covered it in our guide to Copilot visibility for enterprise B2B. The two first-party datasets below are both B2B, and they are the closest public evidence we have.

Finding 6: A few pages take most citations, and the numbers swing

Bing Webmaster Tools launched its AI Performance report as a public preview in February 2026. It shows how often a verified site's pages are cited in Copilot, in AI-generated summaries on Bing, and in some partner integrations, along with the "grounding queries" Copilot used to find them. Two publishers have released their exports.

Search Influence, a marketing agency, pulled 91 days of data (November 12, 2025 to February 10, 2026). It recorded 19,717 Copilot citations across just 86 unique pages. One page, a tool-by-tool comparison of AI SEO tracking software with pricing and named vendors, took 13,599 citations, or 69% of the total. The top four pages took 90%. Average daily citations dropped from 1,520 in the first week of December to 34 in February, a 97% fall in about two months.

Provena, a B2B compliance software company, studied 165 articles over 29 days in August and September 2026. Only 52 articles (32%) were cited at all. The ten most-cited pages took 64% of citations. Category guides that named specific software took 86% of citations, and pages built as "alternatives" or head-to-head comparisons took none.

Those two results look like they contradict each other: a comparison page won almost everything at Search Influence, while comparison pages won nothing at Provena. The likely explanation is the grounding query. Search Influence reports that its grounding queries clustered around tool comparisons, pricing and platform coverage, which is exactly what its winning page offered. Provena's citations clustered on category guides, which suggests its grounding queries were category questions. Copilot does not prefer a format in the abstract. It prefers the page whose structure matches the short query it generated. That is a single-site reading in both cases, and both authors say as much.

The bigger lesson is volatility. A 97% drop in two months means a citation snapshot from last quarter can be badly out of date. Copilot tracking has to be continuous to be useful, which is why Cognizo checks tracked prompts daily rather than on a monthly cycle.

Why Copilot picks these sources: Bing does the retrieval

Microsoft's own documentation explains the first step. According to Microsoft Learn, Microsoft Copilot and Copilot Chat ground their answers by generating search queries and sending them to the Bing search service. Those queries are short, usually a few words built from the user's prompt. Bing returns candidate pages, and Copilot decides which ones to quote.

That two-step process explains most of what the data shows:

• Retail and reference dominate because short, generated queries like "best cordless vacuum price" or "define amortization" retrieve product listings and encyclopedia entries at the top of Bing.
• Low-authority pages get in because selection happens per query. A small page that matches the generated phrase closely can beat a famous page that matches it loosely.
• Format depends on the query, as the Search Influence and Provena exports show.
• If Bing has not indexed you, Copilot cannot cite you. Bing indexing is the entry ticket, not an optional extra for a Google-first team.

The Bing AI Performance report now labels each grounding query by intent (for example informational, commercial or comparison) and topic, and shows a citation share metric for your site. It does not show clicks, it mixes Copilot with other Microsoft surfaces, and it only covers your own site. For a broader explainer on how retrieval and selection work across engines, see how AI search engines find, rank and cite pages.

How Copilot's sources differ from other answer engines

Putting the studies side by side, Copilot has a distinct profile:

The figures in the first two rows come from the same audit, which makes the engine comparison fair. The Reddit and LinkedIn rows come from different studies with different prompt sets, so read them as direction rather than exact gaps.

The upshot is that a win on one engine says little about the others. Compare this profile with our Google AI Overviews citation study and the source mix looks noticeably different. A brand that tracks only ChatGPT can believe it has strong AI visibility while Copilot is quoting a competitor on the same prompt. Cognizo tracks the same prompt across engines side by side so that gap shows up in one view.

What the data means for your Copilot strategy

Each finding points to a specific move. None of them requires guessing.

1. Check Bing first. Confirm your important pages are indexed in Bing Webmaster Tools, submit your sitemap there, and open the AI Performance report. If your best pages are missing from Bing, nothing else on this list matters yet.

2. Find out which of your prompts are "retail," "reference" or "vertical" prompts. A shopping question, a definition question and a vendor-selection question pull from completely different source pools. Group your target prompts by type before deciding what to publish. Cognizo's source classification and page type classification show which kind of page Copilot is citing for each tracked prompt, so the grouping comes from real answers instead of assumptions.

3. Match page format to the grounding query. If Copilot's queries for your topic are comparisons, a clear comparison page with prices and named options is the right asset. If they are category questions, a category guide that names the products in it wins. The Bing report's intent labels and your tracked prompts tell you which.

4. Write the definition better than Wikipedia does. For any term your buyers look up, open the page with a one or two sentence definition, then add what an encyclopedia cannot: numbers, examples from your field, and a direct recommendation.

5. Publish thought leadership as LinkedIn articles, not just posts. The cloro data shows long-form LinkedIn articles have a path into Copilot answers that feed posts and Reddit threads mostly do not.

6. Earn citations on the sources Copilot already trusts in your vertical. In finance, that means Forbes, NerdWallet and specialist review sites. In your category it will be a different list. Pull the domains Copilot cites for your prompts, find the ones that never mention you, and treat them as your outreach and PR targets. Cognizo's competitor tracking in AI search surfaces the domains that cite competitors and not you, which turns this into a concrete list.

7. Spread your bets across pages. When 69% of a site's citations sit on one page, one Copilot change can wipe out most of your presence. Build several strong pages around different grounding queries so a single swing does not take you to zero.

8. Watch it every day. A 97% drop in two months is not unusual in this data. Use a system like Cognizo that re-runs your Copilot prompts daily and shows the day a citation is lost, so you can respond while the change is fresh. The Copilot rank tracking explainer covers how that monitoring works, and our comparison of Copilot rank tracker tools shows how the options differ.

For the page-level work itself, the guide to earning Copilot citations and the Copilot brand optimization guide go step by step.

How Cognizo turns Copilot citation data into work

Public studies tell you what Copilot cites in general. They cannot tell you what it cites for "best compliance software for credit unions" or whatever your buyers actually ask. That is the gap Cognizo fills.

Cognizo runs your tracked prompts through Microsoft Copilot and the other major answer engines every day. For each answer it records which brands are mentioned, which domains and pages are cited, and how your citation share and mention share move over time. Source classification and page type classification show whether Copilot is leaning on retailers, reference sites, editorial coverage or vendor pages for each prompt, and the YouTube and Reddit breakdown shows how much community content is in play.

When a gap appears, Cognizo's content agent can draft and publish an AI-optimized article aimed at that prompt, and technical site audits flag crawlability issues that keep a page out of consideration. Teams that live in other tools can pull the same data through Cognizo's MCP server and API, so a question like "which Copilot prompts did we lose a citation on this week" can be answered from Claude or a reporting stack without opening a dashboard. Regions, languages and competitors are unlimited on every plan, which matters for Copilot because source choices vary by market.

Limitations of this study

We want this to be useful, so here is what it cannot tell you.

• The datasets use different metrics. Ahrefs measures share of citations among the top 50 domains; cloro and LLM Authority Index measure the share of answers citing a domain. We never mix them in one calculation.
• Most of the panel data is US and English. The beauty study is Indonesian. Copilot's sources in Germany or Japan may look different.
• Two datasets are single-site. The Search Influence and Provena exports are precise but describe one site each.
• Some data is older than the rest. Profound's Reddit analysis covers citations through October 2025. We include it because it is the largest cross-engine Reddit dataset available, and cloro's 2026 data points the same way.
• The AI-generated share comes from a detector. Pangram is conservative by the authors' own tests, but no detector is perfect, and the audit used politics, health and environment questions, not commercial prompts.
• Vendors publish most of this research. Ahrefs, cloro, Profound, LLM Authority Index, Search Influence, Obsurfable and Cognizo all sell services related to AI visibility. The arXiv audit is the only fully academic source.

FAQ

What website does Microsoft Copilot cite the most?

Amazon. In Ahrefs' September 2026 data covering more than 3 million US queries, amazon.com held 16.8% of the citation share among Copilot's 50 most-cited domains, followed by walmart.com at 12.6% and en.wikipedia.org at 7.6%. That ranking reflects the large number of shopping questions in the sample. For finance questions, Forbes leads, appearing in 20.73% of Copilot answers in LLM Authority Index's data.

Does Copilot use Bing search results?

Yes. Microsoft's documentation says Copilot generates short search queries from your prompt and sends them to the Bing search service to ground its answer. Copilot then chooses which of the retrieved pages to cite. A page that is not in Bing's index cannot be cited, so Bing Webmaster Tools is the first place to check.

Does Copilot cite Reddit?

Rarely. Profound's cross-engine analysis ranked Reddit 31st among Copilot's cited domains, the lowest of six engines. cloro found Reddit in 0.00% of 55,479 Copilot answers to commercial prompts in summer 2026. Copilot cited LinkedIn, especially LinkedIn articles, more often than Reddit.

How many sources does Copilot cite per answer?

In the 2026 academic audit of 712 real-world questions, Copilot cited an average of 8.44 sources per answer. ChatGPT cited 14.68, Perplexity 7.92 and Gemini 7.26. The number varies a lot with the type of question.

Why does Copilot cite low-quality or AI-generated pages?

Because it selects pages per generated query rather than by domain reputation alone. The same audit found that 27.8% of Copilot's cited sources were likely AI-generated, the highest share of the four engines tested. A page that closely matches Copilot's short search query can be cited even if the site is small or the content is machine-written.

How can I see which of my pages Copilot cites?

For your own site, use the AI Performance report in Bing Webmaster Tools. It shows citation counts per page and the grounding queries behind them, though not clicks or competitors. To see the whole answer, including which competitor pages Copilot cites instead of yours, you need prompt-level tracking. Cognizo records the cited domains and pages for every tracked Copilot prompt, every day.

What kind of content gets cited by Copilot?

It depends on the query Copilot generates. Product listings win shopping queries, encyclopedia-style pages win definitions, and editorial outlets like Forbes and NerdWallet win many finance questions. In first-party Bing data, a detailed comparison page won most citations for one site whose queries were comparisons, while category guides naming specific software won 86% of citations for a site whose queries were about categories.

How often do Copilot citations change?

Often enough that monthly checks miss real changes. One publisher's daily Copilot citations fell from an average of 1,520 to 34 in about two months, a 97% drop, and Ahrefs' monthly top 50 shows domains moving up or down by 10 to 23 places in a single month. Daily tracking is the safest way to catch a loss while you can still respond.

‍