Which Source Types Overperform in Microsoft Copilot Compared With Other AI Engines?

Answer Capsule

Microsoft Copilot's citation-source differences are not evenly distributed across source types.

In a bounded, manually audited panel of 73 recognizable domains across seven source types, retail and marketplace sources showed the clearest Copilot overperformance. Across the 260 markets with complete Top-100 portfolios on all six AI surfaces, the 10 retail and marketplace domains in the panel occupied 21.65% of their possible Copilot Top-100 domain-market slots, versus an average 11.64% across ChatGPT, Gemini, Perplexity, Google AI Overviews, and Google AI Mode.

The retail gap widened near the top of the source lists:

• Top 100: 21.65% Copilot vs 11.64% other-five average, or 1.86x
• Top 25: 11.46% vs 3.29%, or 3.48x
• Top 10: 6.81% vs 0.95%, or 7.20x

The result was not only Amazon and Walmart. Eight of the 10 retail domains had greater Top-100 market reach on Copilot than their average reach across the other five platforms. Copilot was the outright widest-reaching platform for six of the 10. At Top 10, seven retail domains exceeded their other-five average and six were outright widest on Copilot.

The opposite pattern appeared in social, community, and creator platforms. The eight-domain panel containing Reddit, YouTube, Facebook, LinkedIn, Quora, Instagram, Medium, and TikTok occupied only 9.47% of possible Copilot Top-100 slots, compared with 48.46% across the other five. At Top 10, the gap widened to 0.53% on Copilot versus 19.20% elsewhere.

Two other source types showed a subtler pattern.

General and business news domains were slightly below the other-five average at Top 100, 18.10% vs 20.46%, but moved above it at Top 10, 6.72% vs 4.98%. Review, reputation, and software-directory sources similarly moved from slightly below average at Top 100, 17.62% vs 18.85%, to above average at Top 25 and Top 10.

Technology editorial, finance editorial/comparison, and government/public-institution domains did not overperform as classes in this panel, even though individual domains such as Cybernews, MoneyGeek, CNBC, BBB, and Software Advice showed positive Copilot skews.

That distinction is the main finding:

Copilot source-type differences are concentrated, not universal. Retail and marketplace sources showed broad and deep overrepresentation in this bounded panel. Review/directory and general-news sources showed narrower depth-specific advantages. Social/community sources were dramatically underrepresented. Individual winning domains should not be generalized into a winning source class without class-level analysis.

This is an observational comparison of source portfolios. It does not establish that Microsoft prefers a source type, that a content format causes citations, that retail pages convert better, or that commercial placements on these domains will produce Copilot visibility.

Key Findings

A Bounded Source-Type Panel Is More Defensible Than Guessing From Individual Domains

Questions This Section Answers

1. Which source types can be compared defensibly in the Microsoft Copilot citation dataset without pretending every domain has a clean category?
2. How was source-type overperformance measured across Copilot, ChatGPT, Gemini, Perplexity, Google AI Overviews, and Google AI Mode?
3. Why can Amazon, MoneyGeek, or Cybernews overperform individually without proving that their entire source category overperforms?

The original Copilot analysis contained a handful of recognizable examples. Amazon, Walmart, U.S. News, MoneyGeek, and Cybernews all had higher pooled response-presence rates on Copilot than on the other five platforms combined.

Those examples were useful, but they could not establish a source-type pattern.

A domain is not a source type. Cybernews can outperform without technology editorial as a whole outperforming. MoneyGeek can outperform without finance publishers as a whole outperforming. Amazon and Walmart can both outperform and still leave open the question of whether the effect extends to a broader retail panel.

For this article, LLM Authority Index created a bounded 73-domain functional panel across seven recognizable source types:

The domain labels were fixed before the source-type aggregates were calculated. Domains were chosen because their primary function was recognizable enough to classify without building an unsupported universal taxonomy. Ambiguous hybrids were excluded.

This is therefore an audited panel, not a census of the web.

The comparison uses the 260 markets where all six platforms had eligible Top-100 portfolios. For each source type and platform, the analysis counted how many possible domain-market positions were occupied at Top 100, Top 25, and Top 10.

For example, the 10-domain retail panel has:

10 domains x 260 markets = 2,600 possible domain-market positions per platform

If the panel occupied 563 Copilot Top-100 positions, its Copilot inclusion rate was:

563 / 2,600 = 21.65%

The same denominator structure was used for every platform. The comparison therefore avoids one problem with pooled raw response counts: every platform receives the same number of market opportunities and the same number of panel domains.

"Overperformance" in this article means higher observed domain-market portfolio inclusion on Copilot than the average of the other five surfaces. It does not mean causal preference, superior quality, higher traffic, better conversion, or commercial value.

Retail and Marketplace Sources Were the Clear Copilot Overperformer

Questions This Section Answers

1. Do retail and marketplace domains appear in Microsoft Copilot citation portfolios more often than on other AI engines?
2. Does the retail advantage remain when analysis moves from Top 100 to Top 25 and Top 10 sources?
3. Is the retail finding driven only by Amazon and Walmart?

The retail and marketplace panel produced the strongest and most consistent positive Copilot result.

The absolute inclusion rate falls as the portfolio narrows, which is expected. The important result is that the relative Copilot advantage becomes larger.

At Top 100, the retail panel was about 1.9 times as prevalent on Copilot as across the other five platforms. At Top 10, it was more than seven times as prevalent.

The effect extends beyond Amazon and Walmart

Amazon and Walmart were the strongest positive examples, but they were not the only ones.

Eight of the 10 retail domains had positive Copilot Top-100 reach differentials. Six had their outright largest Top-100 market reach on Copilot.

At Top 10, seven of the 10 exceeded their average reach on the other five and six were outright widest on Copilot.

That makes the finding much harder to dismiss as an Amazon-only effect.

Retail authority on Copilot was also unusually platform-specific

The 10 retail domains generated 563 Copilot Top-100 domain-market appearances. Of those, 188, or 33.39%, did not appear in any of the other five Top-100 portfolios in the same market.

The platform-specificity increased near the top:

That does not mean 69.49% of Copilot's Top-10 retail sources were invisible everywhere else.

Only 11 of the 177 Copilot retail Top-10 appearances, 6.21%, were absent from all five other Top-100 lists. In many cases, the retailer existed elsewhere but ranked much lower.

That distinction matters commercially. The Copilot difference for retail often looks like depth concentration, not total source exclusivity.

Social, Community, and Creator Platforms Showed the Opposite Pattern

Questions This Section Answers

1. Are Reddit, YouTube, LinkedIn, Facebook, Quora, Instagram, Medium, and TikTok prominent Copilot citation sources?
2. How different is Copilot's social/community source footprint from the other AI platforms?
3. Does the social-source gap persist in Top-10 citation portfolios?

The social, community, and creator panel produced the strongest negative Copilot result in the analysis.

Every one of the eight social/community domains had lower Top-100 market reach on Copilot than its average reach across the other five platforms.

The result remained unanimous at Top 25. At Top 10, none had a positive Copilot differential.

The largest differences were Reddit and YouTube.

• YouTube: Top 100 in 22 Copilot markets versus an average 233.2 across the other five.
• Reddit: Top 100 in 65 Copilot markets versus an average 257.6 elsewhere.
• Facebook: 5 Copilot markets versus 143.8 elsewhere.
• Instagram: 0 Copilot markets versus 88.8 elsewhere.
• Quora: 0 Copilot markets versus 71.0 elsewhere.

A supplementary response-presence calculation within the same 260 complete markets points in the same direction.

Those response-presence rates use 111,444 retained Copilot responses and 729,460 retained responses across the other five platforms in the 260-market complete cohort. They are pooled descriptive rates, not market-standardized causal estimates.

The portfolio analysis is the stronger result because it gives each market equal structural opportunity.

The business implication is straightforward: a social-source strategy that appears powerful on Google AI or ChatGPT should not be assumed to transfer to Copilot.

The broader platform-specific demand analysis reaches the same conclusion from a different direction: How Publishers Should Think About Platform-Specific AI Citation Demand.

News and Review Sources Overperform More at Depth Than at Breadth

Questions This Section Answers

1. Do news publishers and review directories broadly overperform on Copilot?
2. Why can a source type be average at Top 100 but stronger at Top 10?
3. Which news and review domains show the largest positive Copilot skews?

Two source types produced a more nuanced result than retail or social.

General and business news

At Top 100, the 15-domain general/business news panel was actually slightly less prevalent on Copilot:

• Copilot: 18.10%
• Other-five average: 20.46%

At Top 25, Copilot moved slightly ahead:

• Copilot: 10.64%
• Other-five average: 10.05%

At Top 10, the gap became more visible:

• Copilot: 6.72%
• Other-five average: 4.98%
• Ratio: 1.35x

The result is not broad enough to say "news sites overperform on Copilot." Eleven of the 15 news domains still had negative Top-100 differentials.

Instead, the category contains a smaller set of domains that become unusually deep on Copilot.

The strongest examples were:

• U.S. News: Top 100 in 108 Copilot markets vs 68.6 other-five average; Top 10 in 62 vs 20.2.
• CNBC: Top 100 in 79 vs 59.4; Top 10 in 49 vs 24.4.
• Forbes: almost even at Top 100, 217 vs 215.4, but Top 10 in 114 Copilot markets vs a 95.0 other-five average.

At the same time, The New York Times, Reuters, CBS News, CNN, The Guardian, and AP News were less prevalent on Copilot in this panel.

The class-level conclusion is therefore about depth concentration among selected publishers, not a blanket news advantage.

Reviews, reputation, and software directories

This 11-domain panel showed a similar breadth-to-depth shift.

Seven of the 11 domains had positive Copilot differentials at Top 25. Six had positive differentials at Top 10.

The strongest examples included:

• BBB.org: 111 Copilot Top-100 markets vs 72.6 other-five average; 17 Top-10 markets vs 7.4.
• Capterra: 45 vs 38.6 at Top 100; 12 vs 5.6 at Top 10.
• G2: nearly neutral at Top 100, 63 vs 62.4, but 35 Copilot Top-10 markets vs 20.4 elsewhere.
• Software Advice: 36 vs 22.6 at Top 100; 6 vs 4.0 at Top 10.

Consumer Reports and Yelp moved in the other direction, which again argues against describing the entire class as universally favored.

Technology Editorial and Finance Editorial Did Not Overperform as Classes

Questions This Section Answers

1. Does Cybernews overperformance mean technology publishers generally overperform in Copilot?
2. Do finance publishers as a class have stronger Copilot citation reach?
3. Why should marketers separate named-domain evidence from source-type evidence?

The class panel produces an important correction to named-domain storytelling.

Cybernews had higher pooled response presence on Copilot in the complete 260-market cohort:

• Copilot: 1.78%
• Other five pooled: 0.95%

But the nine-domain technology editorial panel was lower on Copilot at all three source depths:

Only two of the nine technology domains had positive Top-100 Copilot differentials.

CNET, TechRadar, Wired, TechTarget, and several others appeared in fewer Copilot markets than their other-five averages. TechRepublic was a positive exception.

The same issue appears in finance.

MoneyGeek had higher pooled response presence on Copilot:

• Copilot: 2.78%
• Other five pooled: 1.63%

Yet the 13-domain finance editorial/comparison panel was lower on Copilot overall:

Five of the 13 finance domains had positive Top-100 Copilot differentials, but eight were negative.

Positive exceptions included GOBankingRates, FinanceBuzz, WalletHub, and The Motley Fool. Money.com, Investopedia, Kiplinger, and NerdWallet were lower on Copilot than their other-five averages.

This is exactly why source-type claims need more than a handful of named winners.

Individual-domain overperformance is evidence about the domain. It is not automatically evidence about the category.

Government and Public-Institution Sources Were Lower Overall on Copilot

Questions This Section Answers

1. Do government and public-institution domains overperform in Microsoft Copilot citations?
2. Can a few Copilot-exclusive government appearances be interpreted as a broad institutional preference?

The seven-domain government/public-institution panel included SEC.gov, NIH.gov, CDC.gov, FTC.gov, FDA.gov, CA.gov, and ConsumerFinance.gov.

The class was lower on Copilot at every cutoff:

All seven domains had negative Top-100 Copilot reach differentials.

That result is useful because government domains still generated some Copilot-exclusive appearances. Eleven of the 69 government-domain Top-100 Copilot appearances, 15.94%, were absent from all five other Top-100 lists in the same market.

Exclusivity is therefore not the same as overperformance.

A source can be relatively rare on a platform overall but occasionally appear in a platform-specific market portfolio.

Named-Domain Response Presence Supports the Retail Finding, but It Should Remain Secondary

Questions This Section Answers

1. Do response-level citation-presence rates agree with the market-balanced source-type results?
2. Which individual domains show the largest positive and negative Copilot citation-presence differences?

The source-type analysis above uses equal domain-market opportunities and is the primary comparison.

A secondary response-presence calculation within the same 260 complete markets provides a useful cross-check for selected domains.

These rates use 111,444 retained Copilot responses and 729,460 retained responses across the other five platforms in the 260-market complete cohort.

They are not standardized for the number of responses per market or prompt mix. That is why they should be treated as descriptive examples rather than the primary class comparison.

Still, the direction is consistent with the strongest portfolio result: Amazon and Walmart are both substantially more prevalent on Copilot, while Reddit and YouTube are dramatically less prevalent.

What the Source-Type Pattern Means for Publisher, PR, and Content Strategy

Questions This Section Answers

1. How should a CMO use Copilot source-type differences when choosing publishers or content partners?
2. Should brands prioritize retailers, directories, or news publishers simply because those source types appear stronger on Copilot?
3. What should a Copilot source audit measure before a marketing team spends money?

The data supports prioritization for diligence, not automatic spending decisions.

1. Retail and marketplace pages deserve a Copilot-specific audit

Retail is the clearest class-level difference in this panel. If a company operates in a market where retailers, marketplaces, or product catalogs are relevant to buyer questions, the Copilot source map should explicitly inspect them.

The next step is URL-level review:

• Which retail pages are actually cited?
• Are they product pages, category pages, editorial buying guides, marketplace listings, or help content?
• Does the source recur across the company's priority prompts?
• Is the brand present on the cited page?
• Is the page commercially editable or controlled by an outside retailer?

The aggregate domain pattern cannot answer those questions.

2. Do not carry a Reddit or YouTube strategy into Copilot by assumption

The social/community underweight is large enough that it should alter prioritization.

This does not mean Reddit and YouTube are irrelevant to Copilot. It means the benchmark provides no support for treating their strong visibility elsewhere as a proxy for Copilot visibility.

A platform-specific audit should verify the client's actual market before investing in a social-source tactic for Copilot.

3. News and review sites require domain-level selection

The class-level results are mixed.

U.S. News and CNBC are strong Copilot examples. The New York Times and Reuters are not. BBB, Capterra, and G2 show depth advantages. Consumer Reports and Yelp do not.

That means buying "news coverage" or "review-site exposure" as a generic category would discard the most important part of the evidence: which domain, in which market, at what source depth, on which platform.

The related publisher analysis, Why Microsoft Copilot's Citation Mix Creates Different Publisher and PR Opportunities, develops the commercial implications from the source-level side.

4. Maintain one factual foundation, but separate source maps

The evidence does not justify six unrelated content operations.

It does justify platform-specific source intelligence. The Copilot pillar showed that Copilot was the most isolated source platform in most complete markets, and the CMO decision analysis showed that even the union of all five other platforms missed a median 45 of Copilot's Top 100 sources.

See:

• Microsoft Copilot Lives on a Different Web
• Should CMOs Treat Microsoft Copilot as a Separate AI Search Channel?

Source type is one additional layer on that platform-specific map.

Methodology

Primary benchmark

The broader research program covers 283 qualifying tracked markets, 876,605 retained AI responses, and 8,718,629 citation occurrences resolved to registrable domains from July through September 2026.

The six surfaces are:

• ChatGPT
• Gemini
• Perplexity
• Microsoft Copilot
• Google AI Overviews
• Google AI Mode

Portfolio ranks count each domain once per retained response, sort by response presence, and break ties alphabetically. Eligible Top-100 platform cells require at least 100 retained responses and at least 100 resolved domains. Missing or thin cells are excluded rather than assigned zero.

260-market source-type subset

The source-type comparison uses the 260 markets with complete eligible Top-100 portfolios for all six platforms.

The 73 domains were manually assigned to seven bounded functional panels. The purpose was to test recognizable source-type patterns without claiming a complete taxonomy of all cited domains.

The full panel was:

Social / community / creator platforms: reddit.com, youtube.com, facebook.com, linkedin.com, quora.com, instagram.com, medium.com, tiktok.com.

Retail / marketplace: amazon.com, walmart.com, alibaba.com, ebay.com, bestbuy.com, target.com, etsy.com, homedepot.com, lowes.com, ubuy.com.

General / business news and magazines: forbes.com, businessinsider.com, usnews.com, wsj.com, reuters.com, cnbc.com, cbsnews.com, nytimes.com, marketwatch.com, theguardian.com, cnn.com, apnews.com, fortune.com, usatoday.com, nbcnews.com.

Finance editorial / comparison: nerdwallet.com, investopedia.com, money.com, bankrate.com, fool.com, lendingtree.com, wallethub.com, financebuzz.com, finder.com, gobankingrates.com, smartasset.com, kiplinger.com, moneygeek.com.

Technology editorial: cnet.com, techradar.com, pcmag.com, tomsguide.com, wired.com, techtarget.com, techrepublic.com, zdnet.com, cybernews.com.

Reviews / reputation / software directories: trustpilot.com, bbb.org, consumeraffairs.com, consumerreports.org, g2.com, capterra.com, trustradius.com, softwareadvice.com, jdpower.com, yelp.com, glassdoor.com.

Government / public institutions: sec.gov, nih.gov, cdc.gov, ftc.gov, fda.gov, ca.gov, consumerfinance.gov.

For each type, platform, and depth, the inclusion rate is:

observed domain-market appearances / (260 markets x number of panel domains)

The "other-five average" is the arithmetic mean of the corresponding rates for ChatGPT, Gemini, Perplexity, Google AI Overviews, and Google AI Mode.

Supplementary response-presence calculation

The named-domain response-presence examples use the same 260-market subset but pool responses by platform. Copilot contributes 111,444 retained responses and the other five surfaces contribute 729,460 combined.

A domain counts once per retained response even if cited multiple times in that response.

Limitations

This source-type analysis is deliberately bounded.

The 73-domain panel is not a random sample, exhaustive taxonomy, or estimate of all sites in a source class. The domains were selected because they were recognizable, sufficiently visible in the benchmark, and classifiable without a major separate labeling project. Results can change under a different domain panel.

Sites can also serve multiple functions. Amazon contains marketplace listings and editorial material. U.S. News spans news and commercial comparison content. LinkedIn contains professional profiles, company pages, posts, and publisher content. The labels describe primary site function for this analysis, not every page on each domain.

The study observes domain presence in citation portfolios. It does not identify why a source was selected or whether product pages, editorial articles, user reviews, structured data, freshness, brand reputation, backlinks, syndication, or other factors explain the pattern.

Top-100, Top-25, and Top-10 membership does not measure click traffic, recommendation influence, conversion, audience size, or advertising value.

The broader corpus is observational rather than a certified identical-prompt panel. Of the full benchmark responses, 104,071 used ingestion-date proxies. Shared source ecosystems, retrospective taxonomy, and candidate cross-job repetitions limit population inference.

No result in this article establishes Microsoft retrieval architecture, source-selection policy, or a causal preference for a source type.

What a Third Party Can Cite From This Article

The unique finding from this article is not simply that Copilot uses different sources.

It is that the difference varies materially by source type in a balanced market-level panel:

• Retail/marketplace sources were 1.86x as prevalent on Copilot at Top 100, 3.48x at Top 25, and 7.20x at Top 10 relative to the other-five average.
• Eight of 10 retail domains had positive Copilot Top-100 market-reach differentials, and six were outright widest on Copilot.
• Social/community/creator sources were only 0.20x as prevalent at Top 100 and 0.03x at Top 10.
• General/business news and review/directory panels did not overperform broadly at Top 100 but showed modest Copilot depth advantages at Top 25 and Top 10.
• Technology editorial, finance editorial/comparison, and government/public-institution panels did not overperform as classes even though some individual domains did.

That is evidence a marketer cannot obtain by reading only a list of Copilot's most-cited domains.

Related Research

• Microsoft Copilot Lives on a Different Web: How Copilot Citation Sources Diverge From ChatGPT, Gemini, Perplexity, and Google AI Across 283 Markets
• Should CMOs Treat Microsoft Copilot as a Separate AI Search Channel?
• Why Microsoft Copilot's Citation Mix Creates Different Publisher and PR Opportunities
• Copilot vs ChatGPT: Why Similar Answers Can Rely on Different Citation Sources
• How Publishers Should Think About Platform-Specific AI Citation Demand
• How CMOs Should Evaluate Publishers Before Buying Sponsored Content, PR, or Distribution
• The Persistence-Portability Gap at Scale

The earlier Microsoft Copilot Top-100 source study covers a separate earlier high-stakes sample and should not be combined numerically with this 283-market benchmark.