Copilot vs ChatGPT: Why Similar Answers Can Rely on Different Citation Sources

Answer Capsule

ChatGPT and Microsoft Copilot can address the same commercial question while relying on materially different source portfolios. The July through September 2026 benchmark found a 25.00% median Top-100 Jaccard overlap across 268 eligible ChatGPT-Copilot market comparisons.

A stricter 260-market subset, where all six AI platforms had eligible Top-100 portfolios, shows that source divergence continues even after a domain appears on both ChatGPT and Copilot.

In the typical market:

• ChatGPT and Copilot shared 40 of their 100 leading source domains;
• only 4 of ChatGPT's Top 10 sources were also in Copilot's Top 10, and the same was true in the reverse direction;
• 2 of each platform's Top 10 sources were typically absent from the other platform's entire Top 100;
• among domains that did appear in both Top-100 lists, the median market-level Spearman rank correlation was only 0.405;
• the median absolute rank difference among shared domains was 18 positions;
• across 10,017 shared domain-market records, 38.68% differed by at least 25 rank positions and 13.45% differed by at least 50 positions.

The distinction matters because source overlap and source prominence are separate measurements. A domain can exist in both source environments while functioning as a leading authority on one platform and a marginal source on the other.

The title does not mean this study measured semantic similarity between ChatGPT and Copilot answers. It did not. The benchmark measures source portfolios, response presence and source ranks. “Similar answers” describes a possible real-world scenario that requires separate answer-level testing. A paired answer-content study would be needed to quantify whether the two platforms actually reached the same conclusions, recommended the same companies or used similar language.

The evidence supports a narrower conclusion:

A ChatGPT answer audit cannot stand in for a Copilot source audit, even when the answers appear directionally similar to a human reader. Source identity and source prominence need to be measured separately.

This analysis extends Microsoft Copilot Lives on a Different Web, Should CMOs Treat Microsoft Copilot as a Separate AI Search Channel?, and the broader Six AI Engines, Six Different Webs study.

Key Findings

Source Overlap Is Only the First Layer of the ChatGPT-Copilot Difference

Questions This Section Answers

1. How many leading citation sources do ChatGPT and Microsoft Copilot actually share in the same commercial market?
2. Does a 25% Jaccard overlap mean ChatGPT and Copilot share only 25 of their Top 100 sources?
3. Can two AI platforms share a substantial number of domains while still using them very differently?

The first layer is simple set overlap.

Across 268 eligible market-level comparisons, the median ChatGPT-Copilot Top-100 Jaccard overlap was 25.00%.

Jaccard is intersection divided by union. With two equal 100-domain lists, 40 shared domains create a 160-domain union:

40 / 160 = 25.00%

So a 25% Jaccard result does not mean the platforms share only 25 domains. In the strict 260-market complete-platform subset, the median market had exactly 40 shared Top-100 domains.

The distribution was wide:

• first quartile: 34 shared domains;
• median: 40;
• third quartile: 44;
• minimum: 16;
• maximum: 57.

At the low end, Birth Injury Lawyers had only 16 shared Top-100 domains, a Jaccard overlap of 8.70%. DUI Lawyers shared 19. Mesothelioma Lawyers shared 20. Criminal Defense Lawyers and PR Management Agencies each shared 21.

At the high end, Home Builders shared 57 of 100, Dental Insurance shared 54, Mortgage Refinance Lenders shared 53, and Help Desk Software and RV Insurance each shared 52.

That range already shows why there is no single ChatGPT-to-Copilot transfer assumption.

But set overlap still misses a second issue: rank prominence.

Two platforms can agree that a domain belongs somewhere in the Top 100 while disagreeing dramatically about whether it belongs at rank 3, rank 35 or rank 90.

That distinction is especially important for marketers, publishers and researchers. A domain that appears in both lists may look “portable” in a binary presence analysis while playing a very different practical role in each platform's citation environment.

Shared Sources Often Have Very Different Ranks

Questions This Section Answers

1. When ChatGPT and Copilot cite the same domain, do they usually rank that source similarly?
2. How large are the rank differences among sources shared by ChatGPT and Copilot?
3. Is broad portfolio overlap a reliable proxy for agreement about which sources are most important?

The 260 complete markets contained 10,017 domain-market records where the same domain appeared in both the ChatGPT and Copilot Top 100.

The rank agreement was modest.

Across markets, the median Spearman correlation between ChatGPT and Copilot ranks for shared domains was only 0.405. The first quartile was 0.305 and the third quartile was 0.519. Some markets were negative, meaning sources that ranked relatively high on one platform often ranked relatively low on the other.

The median market-level absolute rank difference was 18 positions.

Looking at all 10,017 shared domain-market records directly:

Across all shared records, the median absolute difference was 18 positions, the first quartile was 7 and the third quartile was 35.

This means that simply labeling a domain “shared by ChatGPT and Copilot” can conceal a large prominence difference.

Several examples were extreme:

The important point is not that one platform is “right.” The study does not score source quality or factual accuracy.

The point is that shared presence is not shared prominence.

A source can belong to both evidence environments and still be treated almost oppositely in the rank order.

Portfolio overlap and rank agreement were almost unrelated

One of the more surprising findings is that markets with more shared Top-100 domains did not necessarily rank those shared domains more similarly.

Across the 260 complete markets, the descriptive Pearson correlation between the number of shared Top-100 domains and the Spearman rank agreement of those shared domains was only 0.03.

That is essentially no relationship in this observed sample.

A market can therefore have relatively high domain overlap and weak rank agreement, or low domain overlap but strong ordering among the smaller shared core.

This is why a single cross-platform overlap score is not enough for a source audit.

Top-10 Authority Transfers Poorly Between ChatGPT and Copilot

Questions This Section Answers

1. If a domain is a Top-10 citation source on ChatGPT, how often is it also Top 10 on Copilot?
2. How often does a leading Copilot source disappear from ChatGPT's Top 100 entirely?
3. Does Top-25 source authority transfer more reliably than Top-10 authority?

Top-10 transfer is more commercially intuitive than a full Top-100 overlap score.

Across 260 complete markets, ChatGPT contributes 2,600 Top-10 domain positions and Copilot contributes another 2,600.

Only 993 ChatGPT Top-10 positions, 38.19%, were also Top 10 on Copilot.

The rest were distributed as follows:

The reverse direction was almost equally weak.

So roughly one-fifth of each platform's Top-10 source positions did not appear anywhere in the other platform's Top 100.

In the median market:

• 4 of ChatGPT's Top 10 were also Copilot Top 10;
• 6 reached Copilot's Top 25;
• 7 reached Copilot's Top 50;
• 8 appeared somewhere in Copilot's Top 100;
• 2 were missing from Copilot's Top 100.

The reverse medians were nearly identical.

Only 3 of 260 markets had zero mutually shared Top-10 domains. But weak mutual Top-10 overlap was common:

• 16 markets had one or fewer;
• 61 markets, 23.46%, had two or fewer;
• 104 markets, 40.00%, had three or fewer;
• only 93 markets, 35.77%, had five or more;
• only 8 markets had seven or more;
• one market reached eight.

Top-25 authority does not solve the transfer problem

The same pattern persists when the portfolio is expanded to Top 25.

Across 6,500 ChatGPT Top-25 positions:

• 40.85% were also Copilot Top 25;
• 67.71% appeared somewhere in Copilot's Top 100;
• 32.29% were absent from Copilot's Top 100.

In the reverse direction:

• 40.85% of Copilot Top-25 positions were also ChatGPT Top 25;
• 67.31% appeared somewhere in ChatGPT's Top 100;
• 32.69% were absent from ChatGPT's Top 100.

The symmetry is notable. Neither platform looks like a simple broader or narrower version of the other.

The two source systems are selecting different leading domains in both directions.

Familiar Domains Split Sharply Between the Two Platforms

Questions This Section Answers

1. Which major websites were frequently Top-10 sources on ChatGPT but not on Copilot?
2. Which major websites were frequently Top-10 sources on Copilot but weak or absent on ChatGPT?
3. Are there any domains that transfer well between ChatGPT and Copilot?

The aggregate transfer pattern becomes easier to understand through recognizable domains.

ChatGPT-heavy Top-10 sources

Reddit is the clearest example. It was a ChatGPT Top-10 source in 249 complete-market observations but a Copilot Top-10 source in only two of those markets. In 184 of the 249, Reddit was absent from Copilot's Top 100 entirely.

When Reddit did appear in Copilot's Top 100 for those ChatGPT-leading markets, its median Copilot rank was 67.

But not every source was platform-skewed.

Forbes transferred much better. Of its 165 ChatGPT Top-10 appearances, 110 were also Copilot Top 10 and only six missed Copilot's Top 100. NerdWallet was even more portable near the top: 61 of 71 ChatGPT Top-10 appearances were also Copilot Top 10.

G2, ConsumerAffairs, Bankrate, Insurance.com and MoneyGeek also showed relatively strong transfer in the markets where they reached ChatGPT's Top 10.

That contrast is important. “ChatGPT source” and “Copilot source” are not fixed domain labels. Portability differs by publisher and category.

Copilot-heavy Top-10 sources

The opposite side is equally striking.

Amazon appeared in Copilot's Top 10 in 76 complete-market observations. It was not Top 10 on ChatGPT in any of those 76 markets, and it was absent from ChatGPT's entire Top 100 in 58.

U.S. News showed a similar pattern: 62 Copilot Top-10 appearances, zero corresponding ChatGPT Top-10 appearances and 41 complete absences from ChatGPT's Top 100.

Walmart reached Copilot's Top 10 in 61 markets but ChatGPT's Top 10 in only two of those cases.

CNBC reached Copilot's Top 10 in 49 observed markets and ChatGPT's Top 10 in none of those same market positions.

PCMag was even more discrete in this directional test. All 20 observed Copilot Top-10 cases were absent from ChatGPT's Top 100.

MSN had 11 Copilot Top-10 market appearances and was absent from ChatGPT's Top 100 in all 11.

Again, this does not prove platform preference, source quality or commercial impact. It shows that leading-source prominence can be radically platform-specific even for familiar domains.

High Portfolio Overlap Can Still Hide Weak Top-10 Agreement

Questions This Section Answers

1. Can ChatGPT and Copilot share half of their Top-100 source lists but still disagree about the most prominent sources?
2. Are portfolio overlap and Top-10 agreement interchangeable metrics?

Several markets demonstrate why set overlap and rank depth need to be reported separately.

Help Desk Software: 52 shared Top-100 domains, only two shared Top-10 domains

Help Desk Software had 52 shared Top-100 domains, producing a relatively high 35.14% Jaccard overlap.

Yet only two domains were Top 10 on both platforms.

The shared-domain rank correlation was only 0.302, and the median absolute rank difference among shared sources was 27 positions.

ChatGPT's Top 10 included TechRadar at #1, TechTarget at #8 and Reddit at #9. All three were absent from Copilot's Top 100.

Copilot's Top 10 included Help Scout at #4, which ranked #48 on ChatGPT, and The CX Lead at #6, which ranked #93 on ChatGPT.

G2 and Atlassian were among the few genuinely aligned sources, ranking #3 on both and #7 on both, respectively.

A binary overlap metric would describe this market as relatively shared. A prominence analysis shows much deeper disagreement.

Background Checks: 49 shared domains, only one mutual Top-10 source

Background Checks shared 49 of 100 domains, a 32.45% Jaccard overlap.

Yet ChatGPT and Copilot had only one domain in common across their Top 10s.

This is exactly the type of market where a CMO could mistakenly conclude that the platforms share a similar publisher environment because nearly half of the Top-100 source list overlaps.

The leading-source environment tells a different story.

Customer Service Software: 50 shared domains, two mutual Top-10 sources

Customer Service Software shared 50 domains and had a 33.33% Top-100 Jaccard overlap. Only two sources appeared in both Top 10s.

That pattern reinforces a general measurement rule:

Top-100 overlap measures ecosystem similarity. Top-10 transfer measures prominence similarity. They are not substitutes.

Low Portfolio Overlap Can Still Contain a Stable Shared Core

The reverse pattern also occurs.

Job Posting Sites: only 25 shared Top-100 domains, but strong rank agreement within the shared core

Job Posting Sites shared only 25 Top-100 domains, a 14.29% Jaccard overlap.

But among those 25 shared domains, the rank Spearman correlation was 0.814, the highest in the 260-market complete subset. The median absolute rank difference was only 7 positions.

Five domains were Top 10 on both platforms:

• LinkedIn;
• Indeed;
• Glassdoor;
• ZipRecruiter;
• Glassdoor Australia.

The top of the two portfolios still diverged in meaningful ways. TechRadar and Business Insider were ChatGPT Top-10 sources but absent from Copilot's Top 100. SimplyHired, CareerBuilder and Analytics Insight were Copilot Top-10 sources but absent from ChatGPT's Top 100.

Still, the shared core was unusually coherent.

This example is important because it prevents an oversimplified conclusion that low portfolio overlap always means weak agreement everywhere.

A small shared core can be tightly aligned even while the surrounding source universes diverge.

Bankruptcy Lawyers: 24 shared domains, seven mutual Top-10 sources

Bankruptcy Lawyers shared only 24 Top-100 domains, yet seven domains were common to both Top 10s.

This is another example of deep agreement within a relatively small shared set.

The broader implication is that source breadth and source depth must be measured together.

Budgeting Apps Shows How Rank Order Can Nearly Invert

Questions This Section Answers

1. Can ChatGPT and Copilot share many of the same domains but rank them in almost unrelated order?
2. What does a near-zero cross-platform rank correlation look like in a real market?

Budgeting Apps is one of the clearest examples.

ChatGPT and Copilot shared 42 Top-100 domains, producing a 26.58% Jaccard overlap. That is slightly above the overall ChatGPT-Copilot median.

But the Spearman rank correlation among those shared domains was -0.029, effectively zero. The median absolute rank difference was 33 positions.

Some of the shared leaders were aligned:

• Forbes ranked #1 on ChatGPT and #2 on Copilot;
• NerdWallet ranked #2 and #4;
• The Penny Hoarder ranked #6 and #1;
• WalletHub ranked #5 and #10.

But the rest of the list diverged sharply.

ChatGPT ranked Reddit #4; Copilot ranked it #53.

ChatGPT ranked WTOP #9; Copilot ranked it #97.

Copilot ranked CNBC #3; ChatGPT ranked it #63.

Copilot ranked WorldMetrics #5; ChatGPT ranked it #69.

PCMag and LendEDU were Copilot Top-10 sources but absent from ChatGPT's Top 100.

TechRadar and Kiplinger were ChatGPT Top-10 sources but absent from Copilot's Top 100.

The result is a market where the platforms share a reasonable number of domains yet organize most of that shared evidence very differently.

For marketers, the lesson is straightforward: presence on both platforms does not tell you whether a source is strategically prominent on both.

Dental Implants Shows Full Source-Environment Divergence

Dental Implants sits near the opposite extreme.

The platforms shared only 23 Top-100 domains, a 12.99% Jaccard overlap. Their shared-domain rank correlation was 0.301 and the median absolute rank difference was 25 positions.

Only two sources were Top 10 on both platforms: ClearChoice and BBB.

Four ChatGPT Top-10 sources were absent from Copilot's Top 100:

• Reddit;
• RealSelf;
• Forbes;
• NIH.

Five Copilot Top-10 sources were absent from ChatGPT's Top 100:

• Nuvia Smiles;
• Real Dental Costs;
• The Pricer;
• How Much Blog;
• 4M Dental Implants.

Dental Implants therefore demonstrates all three forms of divergence at once:

1. low set overlap;
2. weak rank agreement among the shared sources;
3. substantial leading-source exclusivity in both directions.

A screenshot comparison of two final answers would not reveal any of those source-portfolio differences unless the citations themselves were inspected.

Industry Context Changes ChatGPT-Copilot Transfer

Questions This Section Answers

1. Are ChatGPT and Copilot citation sources more transferable in finance than in legal or health markets?
2. Which industries show the weakest ChatGPT-Copilot source overlap?

The 260 complete markets show clear descriptive sector differences.

Legal & Compliance was the least portable major macro sector by shared-domain count and Jaccard overlap. The median legal market shared only 25 Top-100 domains between ChatGPT and Copilot and only three mutual Top-10 sources.

Finance & Insurance was more transferable, with a median 43 shared Top-100 domains and five mutual Top-10 sources.

But even finance is far from a common source map. A median 27.39% Jaccard still leaves most of the union nonshared.

Sector averages should therefore guide expectations, not replace market-level analysis.

Similar-Looking Answers Need a Three-Layer Audit

Questions This Section Answers

1. How should a CMO compare ChatGPT and Copilot if both appear to recommend similar companies?
2. What should an AI visibility report measure beyond final-answer text?
3. How can marketers distinguish recommendation agreement from source agreement?

The study does not measure semantic answer similarity. That limitation is central to how the findings should be used.

If a future monitoring program wants to know whether ChatGPT and Copilot produce similar answers, it should collect and compare at least three separate layers.

Layer 1: Answer and recommendation agreement

Record what the user-facing response actually says:

• companies recommended;
• company ordering where applicable;
• factual claims;
• sentiment;
• reasons for the recommendation;
• exclusions or warnings.

This is the layer most marketers notice first.

Two platforms can look similar here because they mention the same companies or reach similar conclusions.

Layer 2: Citation-source agreement

Separately record:

• cited domains;
• cited URLs;
• source categories;
• owned versus third-party sources where ownership is verified;
• source recurrence across responses.

This is where the ChatGPT-Copilot divergence in this study becomes visible.

A similar recommendation can be supported by very different evidence environments.

Layer 3: Citation prominence and page-level evidence

Finally, measure:

• source rank or response presence;
• whether the source sits in the leading Top 10, broader Top 25 or tail of the Top 100;
• which specific pages are being cited;
• whether the cited pages support the same claim;
• whether the source persists over time.

The rank analysis in this article shows why this third layer matters. Even when the same domain appears on both platforms, it may carry very different prominence.

The AI Citation Authority vs AI Recommendation Authority article develops the broader distinction between being cited and being recommended.

What This Means for Publisher, PR and Content Strategy

The source differences create practical consequences, but the evidence should not be stretched beyond what it supports.

1. Do not buy “ChatGPT authority” and assume Copilot transfer

A publisher that performs strongly on ChatGPT may be peripheral or absent on Copilot in the same market.

Reddit, TechRadar, Consumer Reports, TechTarget and Reuters provide strong examples of ChatGPT-heavy prominence in this dataset.

Conversely, Amazon, U.S. News, Walmart, CNBC, PCMag and MSN often held much stronger Copilot positions.

A commercial partner should therefore document platform-specific category evidence rather than present one generic AI citation claim.

2. Do not treat shared domains as equally valuable on both platforms

A domain present in both Top-100 lists can still differ by 50, 75 or even nearly 100 positions.

For an advertiser, that changes the diligence question from:

“Is this publisher cited by both platforms?”

To:

“How prominent is this publisher in my market on each platform, and which cited pages support that prominence?”

3. Keep source strategy coordinated, not duplicated

The findings do not imply that every content program needs two separate teams.

Shared company facts, original data, product documentation and strong category content can remain common infrastructure.

What needs separation is the measurement and source-opportunity layer:

• which third-party domains matter;
• which pages are cited;
• which publisher relationships are relevant;
• which platform-specific gaps exist.

The Copilot publisher and PR opportunity analysis shows how to turn that source distinction into a bounded opportunity map.

4. Measure outcomes after action

The dataset does not establish that earning coverage from a Copilot-heavy source will cause Copilot to recommend a brand.

It also does not establish that a ChatGPT-heavy publisher is ineffective on Copilot.

Any sponsorship, editorial outreach, factual correction or owned-content change should be monitored as an intervention with separate outcome measures.

Citation changes, recommendation changes, referral traffic and business performance are different outcomes.

Methodology

This article uses the frozen July through September 2026 LLM Authority Index research corpus.

The broader benchmark contains:

• 283 qualifying tracked markets;
• 876,605 retained AI responses;
• 8,718,755 citation-array slots;
• 126 unresolved slots;
• 8,718,629 citation occurrences resolved to registrable domains;
• 6,825,821 response-domain appearances;
• 308,807 distinct registrable domains.

The six measured surfaces were:

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

The aggregate ChatGPT-Copilot Top-100 Jaccard median uses 268 eligible market-level pair comparisons.

The article-specific rank-transfer analysis uses the stricter 260-market subset where all six platform portfolios were eligible with at least 100 responses and at least 100 resolved domains for the frozen July through September scope.

For each market-platform cell, domains were ranked by response presence, counting each domain once per retained response. Ties were broken alphabetically. Top-10, Top-25, Top-50 and Top-100 cutoffs were then compared directly.

Shared-source rank agreement uses Spearman correlation within each market across domains appearing in both ChatGPT's and Copilot's Top 100. Absolute rank differences are calculated on the same shared-domain records.

Top-10 transfer asks where a source ranked on the other platform:

• Top 10;
• 11-25;
• 26-50;
• 51-100;
• or absent from the Top 100.

The same procedure was repeated in both directions.

Limits

This is descriptive observational research, not a controlled platform experiment.

The study does not establish:

• semantic similarity between ChatGPT and Copilot answers;
• identical prompts or stable matched response panels across every platform and period;
• causal source-selection mechanisms;
• source quality or factual accuracy;
• platform architecture;
• user traffic or market share;
• recommendation influence;
• conversion or revenue effects;
• future ranking behavior.

Of the retained responses, 104,071 used ingestion-date proxies. The frozen corpus does not support a certified identical-prompt, identical-model, identical-time comparison across the full benchmark.

Only 246 markets had all three months and all six platforms eligible in the marginal portfolio analysis used elsewhere in the corpus. The 260-market subset here is a different requirement: six eligible platform Top-100 portfolios in the pooled July through September scope.

Domain-level comparison also hides URL-level differences. ChatGPT and Copilot can cite the same registrable domain while using different pages, documents or article types.

The analysis therefore supports source-environment comparison, not claims about how either platform technically retrieves information.

Conclusion

ChatGPT was Microsoft Copilot's nearest citation-source neighbor in most markets in the broader Copilot analysis. But “nearest” is not the same as “interchangeable.”

Across the strict 260-market comparison:

• the median market shared only 40 Top-100 domains;
• shared-source rank correlation was only 0.405;
• the median shared source differed by 18 rank positions;
• 38.68% of shared records differed by at least 25 positions;
• only 38.19% of one platform's Top-10 source positions were also Top 10 on the other;
• roughly one in five Top-10 sources disappeared from the other platform's Top 100 entirely;
• roughly one in three Top-25 sources disappeared from the other Top 100.

The most important insight is therefore not merely that ChatGPT and Copilot cite different domains.

It is that they can disagree at every layer of source authority: which domains appear, which domains lead, and how strongly shared domains are ranked.

If the final answers look similar, that similarity should be measured separately rather than assumed to imply common evidence.

For AI visibility teams, the operating model is straightforward:

1. compare answer and recommendation outcomes;
2. compare cited domains and URLs;
3. compare source prominence;
4. keep platform-specific gaps visible inside one coordinated reporting system.

That is the difference between measuring what an AI answer says and measuring the evidence environment that sits behind it.

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?
• Which Source Types Overperform in Microsoft Copilot Compared With Other AI Engines?
• Why Microsoft Copilot's Citation Mix Creates Different Publisher and PR Opportunities
• Six AI Engines, Six Different Webs: Comparing Citation Source Overlap Across ChatGPT, Gemini, Perplexity, Copilot, Google AI Overviews, and Google AI Mode
• Can One GEO Strategy Work Across ChatGPT, Gemini, Perplexity, Copilot, Google AI Overviews, and Google AI Mode?
• AI Citation Authority vs AI Recommendation Authority: Why Being a Source Is Not the Same as Being Recommended
• The Persistence-Portability Gap at Scale: 282 of 283 Markets Show Higher AI Citation Persistence Than Cross-Platform Portability Across 8.72 Million Citation Occurrences

The earlier Persistence-Portability framework provides conceptual context. The 2026 AI Citation Authority Study covers an earlier high-stakes sample and is a separate population.