ChatGPT vs Perplexity Citations: What Our Repeat Runs Show, and How to Measure Each Engine
Short version. On the same customer question, ChatGPT and Perplexity cited largely different sites in our 40-call study on 10 November 2026. On the listed layer, with lists cut to a common length, they shared 0.057 of their cited sites, while ChatGPT shared 0.720 with its own same-day re-run. A gap that size cannot be put down to same-day run-to-run wobble, so cross-engine overlap only means something when it is read against each engine's agreement with itself. We therefore count each engine separately, with repeat runs, one named citation layer and size-matched lists.
Do ChatGPT and Perplexity cite the same sources for the same question?
Mostly not. Qwairy, which sells AI-search monitoring, put the pair at about 7% of cited domains per question in its 13 March to 11 June 2026 data [1]. Wellows, which sells AI-visibility tools, found that Perplexity never touches 89.1% of the websites ChatGPT cites (January to June 2026) [2]. Against a random baseline, though, Wellows ranks this pair the second most alike of the ten it compared [2]. Without size matching, our own study found 6.0% of cited domains per question and 1.4% of exact pages in common (10 questions, 10 November 2026).
"Overlap" here means cited-source overlap: of all the sites either answer cites, the share that both cite (the Jaccard index, where 1 means identical sites and 0 means nothing in common). A "site" is the registrable domain, such as nhs.uk.
Why the two engines retrieve differently is covered in ChatGPT vs Perplexity: How Each Engine Finds Your Content.
Is the gap real, or is each engine just disagreeing with itself?
In our study the gap was real. On the listed layer, with each question's lists cut to a common length, ChatGPT and Perplexity shared 0.057 of their cited sites, against 0.720 for ChatGPT with its own same-day re-run.
What we observed
The gap was 0.66 (97.5% CI 0.47 to 0.87). We took it question by question against the lower of the two engines' self-agreement, which on the listed layer was ChatGPT's every time. We then averaged over 5 clusters, each pairing a question's UK and US versions. It was positive in all 5 clusters and in all 10 questions (sign test over clusters, one-sided p = 0.031; post-hoc check). Before collecting data we set 0.35 of self-agreement as the highest cross-engine figure that would still match Qwairy's pattern, and ours came in well below it.
The same pattern held for the sources each answer's text links to. On that layer, cross-engine overlap was 0.050, against 0.727 for ChatGPT with itself and 0.600 for Perplexity with itself. The gap was 0.51 (97.5% CI 0.24 to 0.83), taken question by question against the lower self-agreement on each question, so it differs from the subtraction of the printed means. The linked lists were short, between 1 and 4 sites after size matching, which makes each question's figure move in large steps.
Scroll to see the full figure
Two checks on whether one site drove the result
On the UK emergency-dentist question, all ten of Perplexity's listed sources were pages on nhs.uk, so that question's size-matched comparison had only one site to compare. Removing the question leaves 9, and the finding holds. Taking reddit.com out of every list also left the comparison unchanged.
How our run compares with Qwairy's repeat-run figures
Qwairy puts ChatGPT's agreement with itself at roughly 28% and Perplexity's at roughly 41% in its 13 March to 11 June 2026 data, from re-runs a few days apart [1]. Qwairy does not state which citation layer it counts, and our repeats were minutes apart on one day, so the comparison is loose.
On the listed layer, Perplexity's self-agreement minus ChatGPT's was +0.28 across our 10 questions (99% CI +0.04 to +0.49). That is the same direction as Qwairy's March to June 2026 figures, but inflated, because Perplexity returned an identical source list on both runs. On the linked layer, over the same 10 questions, it was -0.13, the opposite direction. The interval there is unstable at 5 clusters: its 99% lower bound is about -0.3 to -0.4 depending on the resampling method, and its upper bound is 0.00. The run is too small to settle it.
Why do published overlap figures for this pair run from about 6% to 13%?
Each figure answers a slightly different question. The table sets out what each publisher states, with our own unmatched figure beside them.
| Source | Data stated | Unit and metric | ChatGPT and Perplexity |
|---|---|---|---|
| Qwairy (sells AI-search monitoring) [1] | 13 March to 11 June 2026; client brands; each engine's most recent answer per question | Registrable domain; Jaccard per question; monitored brand's own domain excluded | About 7% (13 March to 11 June 2026) |
| Wellows (sells AI-visibility tools) [2] | January to June 2026; each question collected once per engine; English only, 27 markets; a pair uses every question that pair answered | Agreement on websites cited (metric not named on the page) | 6.5% (January to June 2026) |
| Writesonic (sells AI-visibility tracking) [3] | Collection window not stated; 119,513 prompts for this pair | Domain level, URLs lowercased and "www." stripped; Jaccard per prompt; pair counted only where both engines returned citations | 13% (window not stated) |
| Our study | 10 November 2026; 10 questions, two runs per engine | Registrable domain; Jaccard per question, no size matching (descriptive comparator) | 6.0% of domains, 1.4% of exact pages |
Scroll to see the full figure
Where the studies differ:
- Grain. Counting exact pages gives a lower figure than counting sites. Qwairy reports about 3% at exact-URL level in its 13 March to 11 June 2026 data [1], and Writesonic notes that exact-URL matching would push every score lower [3]. For the pair in our run of 10 questions, exact-page overlap came to 0.019 after size matching (directional only).
- Pair base. Writesonic counts the pair only on prompts where both engines returned citations [3]. A study that keeps the other questions is counting a different set of prompts.
- The brand's own domain. Qwairy removes the monitored brand's domain in its March to June 2026 data [1], and Writesonic does not say whether it did [3]. A brand's own site can be among the cited sources, so keeping it in or taking it out can move the figure.
- One answer or several. Wellows collected each question once per engine and says that run-to-run variation is unmeasured [2]. A source cited only once is bound to look engine-specific, and the citation index that Machine Relations Research publishes shows how large that group is. In it, 46.2% of 22,377 cited domains were cited exactly once (10 May to 21 September 2026, six engines), and at most 67.4% were seen on only one engine [5]. Among domains cited at least ten times, at least 98.5% were cited by two or more engines [5]. Only its ChatGPT, Claude and Perplexity arms ran without a collection gap, so we use it for the single-citation effect and take no figure from it for this pair.
Wellows' chance-adjusted figure for the pair is 19.1 times above chance [2]. Wellows reads ChatGPT's raw gap as a difference in the slice of the web it draws from [2]. A chance comparison belongs beside any overlap figure, and step 8 below shows the one we used.
The overlap figure also depends on whether you ask how much two lists share or how much of one list is in the other. Take two answers, one citing six sites and the other four, with two sites in common. Of the eight sites cited in all, two are shared, an overlap of 0.25. Two of the shorter list's four sites appear in the longer one (50%), while two of the longer list's six appear in the shorter (33%). Wellows' 89.1% and 90.1% (January to June 2026) are one-way figures of this second kind: the share of one engine's websites that the other never cites [2]. Qwairy (13 March to 11 June 2026) and Writesonic report Jaccard [1][3].
Which citations do you count: everything an engine lists, or what the answer links to?
Count both, in separate columns. The listed layer is every source the API returns for the answer. The inline layer is what the answer text itself points at: ChatGPT's in-text links, and Perplexity's sources that carry a numbered marker in the text. Which engine appears to cite more depends on the layer.
In our run of 20 answer pairs, ChatGPT's API listed a median of 17 sites per answer against 9 for Perplexity (ranges 11 to 21 and 1 to 10). Across all 20 answers that came to 587 listed URLs for ChatGPT and 200 for Perplexity. The sites linked in the answer text were close in number and slightly higher for Perplexity: a median of 3 (range 1 to 5) for ChatGPT against 4.5 (range 1 to 7). The layer gap was 10.0 sites (95% CI 9.0 to 11.0), a difference of differences computed on per-answer means, so it does not follow from the medians printed here. This is directional, from a simple mean difference with a cluster bootstrap.
Scroll to see the full figure
RankSurf, which sells AI-visibility tracking, counted third-party citation occurrences instead. On the same 559 prompts, in answers dated 23 April to 18 August 2026, Perplexity produced 3.0 times as many as ChatGPT [4]. Our listed counts lean the other way and our linked counts are close, so the ordering depends on the layer and on how many sources each API returns.
On that last point, Perplexity returned exactly 10 listed URLs on each of our 20 answers, consistent with a fixed return length in the sonar-pro API set-up we used. The figure describes our 10 November 2026 run through that API, not an engine property. Only 99 of the 200 URLs Perplexity listed (49.5%) were marked in its answer text (95% interval 42.6% to 56.4%, a range that treats URLs as separate observations although they cluster by answer).
A list can repeat while the answer does not. Perplexity returned the same 10-source list, in the same order, on both runs of all 10 questions, yet its answer text differed between runs (text similarity between 0.03 and 0.88). ChatGPT's source list differed between runs on all 10 questions. The repeated list shows that retrieval repeated within minutes. It says nothing about whether the answer was stable.
A site can dominate one layer and barely register in the other. In our run Reddit was listed in all 20 of ChatGPT's answers (95% interval 84% to 100%), where it made up 195 of 587 listed URLs. The answer text linked it in only 2 of the 20. Its first position in each ChatGPT list was always 13th to 20th by listed URL (9th to 19th by distinct domain), beyond the size-matched cut. Perplexity listed Reddit in none of its 20 answers. Those are counts for ten local-service questions on one day, and RankSurf reports Reddit leading Perplexity's third-party sources in its own sample [4]. A listed-layer count would have called Reddit a major ChatGPT citation, whereas an inline count would not.
Can Perplexity cite your website without naming your business?
Yes, and ChatGPT can too, so "cited" and "named" are separate outcomes to record. RankSurf's matched answers cover 49 tracked brands, dated 23 April to 18 August 2026, and are a convenience sample of its own projects. In its discovery answers Perplexity cited a tracked brand's site without naming the brand in 27 of its 80 own-site-cited answers, or 33.8%, and ChatGPT did the same in 5 of 31, or 16.1% [4].
We did not measure this in our own run. Record the two states in separate columns: an answer that links your site without naming you, and an answer that names you while citing someone else.
How do you measure citations on each engine so the comparison holds up?
Run one fixed set of questions on both engines and ask every question more than once on the same day. Record the cited sites from one named layer, cut each group of lists to the same length, and calculate overlap between the engines and within each engine. Read the between-engine figure against the within-engine ones, and against a chance comparison. The steps follow what we ran, with two differences: we ran ten questions through an API, whereas the steps allow manual collection in each app and suggest more questions.
- Write the question set. Use questions a buyer would type, none naming your business. Questions that do name you get their own set and their own tally. Run as many questions as you can twice on each engine: our 10 gave coarse intervals, particularly once split by sector.
- Fix the conditions and change only the engine. Same wording, same day, same location setting (we set the search country to each question's market), and web search switched on for ChatGPT. Write down the model each run reports. Ours were gpt-5.5 (returned as gpt-5.5-2026-04-23) and sonar-pro, both through the DataForSEO LLM Responses API, which is not the same thing as the consumer apps.
- Ask every question at least twice per engine, minutes apart. The second run gives you an engine's agreement with itself. Runs a few days apart may drift further, so record which you did. How many repeats a given question needs depends on how far its answers move, which Can You A/B Test Content for AI Visibility? What You Can Run Instead works through.
- Save two lists for every answer, one for the listed layer and one for the inline layer. Keep the raw response, the engine, the run number and the time. Flag empty or truncated answers. Reconcile your totals, so that calls planned, raw files saved and rows in your table are the same number (ours were 40 each).
- Normalise before you compare. Lowercase everything, strip "www." and tracking parameters, then reduce each URL to its registrable domain for site grain and keep the full path for page grain. Decide whether to exclude your own domain, apply the rule to both engines, and write it down.
- Match list sizes. Within each question's group of answers, cut every list to the length of the shortest, keeping each engine's own order. Without this step, an engine that returns longer lists looks as if it disagrees more, when it has only returned more. If ChatGPT lists 17 sites and Perplexity lists 9, for example, keep ChatGPT's first 9.
- Calculate three overlaps per question and layer. Overlap is shared sites divided by the sites in either list. Take ChatGPT's run 1 against its run 2, Perplexity's run 1 against its run 2, and ChatGPT against Perplexity across all four run pairings, then average. Average each figure across the questions in each sector or market and report how many questions went in. Compare the cross-engine figure with the lower of the two self-agreement figures, which is the stricter baseline for the claim you are testing.
- Add a chance comparison. Pair each answer with the other engine's answer to a different question in the same market, and calculate the same overlap. If same-question overlap does not stand clearly above different-question overlap, both engines may be drawing on one market-wide pool of sites.
- Report the layer, the grain, n and the caveat. Counts are lower bounds, because a source an engine used but did not return cannot be counted. Check whether one site dominates a question: we re-ran the headline comparison without nhs.uk and again without reddit.com. Present per-sector and per-market figures as directional unless each holds enough questions to carry a conclusion.
Here is the arithmetic on invented lists, so the steps can be checked by hand. After size matching to four sites, ChatGPT's two runs cite A, B, C, D and A, B, C, E, while Perplexity's two runs cite A, F, G, H and A, F, G, I.
| Comparison | Shared sites | Sites in either list | Overlap |
|---|---|---|---|
| ChatGPT run 1 against run 2 | 3 (A, B, C) | 5 | 0.60 |
| Perplexity run 1 against run 2 | 3 (A, F, G) | 5 | 0.60 |
| ChatGPT against Perplexity, mean of four pairings | 1 (A) | 7 | 0.14 |
The cross-engine figure of 0.14 is about 0.24 of the self-agreement figure of 0.60.
On the chance comparison, our run (40 question pairs from the first of the two runs, descriptive only) gave 0.033 across engines for different questions in the same market, against 0.057 for the same question. The pre-registered comparator could not be computed with this design, so this is a recorded deviation. The engines shared nearly as few sites on different questions as on the same one. Wellows' chance adjustment works differently, correcting for how much each pair's candidate pools overlap [2], so the two figures do not compare.
What did our study observe across UK and US sectors?
On 10 November 2026, from 14:30 to 14:38 UTC, we ran 10 customer-style questions through the DataForSEO LLM Responses API: dental, home trades and legal services, in the UK and the US. Six were city questions: an emergency dentist, roofers and a car-accident injury lawyer, each asked about one UK city and one US city. The other four were two identical-text questions (implant cost and rewiring cost), each asked once with UK and once with US search settings. Each question ran twice on ChatGPT (gpt-5.5) and twice on Perplexity (sonar-pro), with the search country set to the question's market. All 40 calls returned a result, with no retries, exclusions or re-runs.
These are results from a small study on one day, not a market benchmark. The question set is ours, the models are API versions rather than the consumer apps, and newer ChatGPT models exist. Only the pooled comparison set out above is stated as a result. Everything below is what we observed, and it points in a direction without settling it.
| Sector | Questions | Cross-engine overlap | ChatGPT agreement with itself |
|---|---|---|---|
| Dental | 4 | 0.015 | 0.72 |
| Home trades | 4 | 0.078 | 0.70 |
| Legal | 2 | 0.10 | 0.76 |
Scroll to see the full figure
Split by market, the listed-layer figures were 0.103 against 0.686 in the UK (5 questions) and 0.012 against 0.753 in the US (5 questions). In our smaller run the gap held in every sector and in both markets, though with 2 to 4 questions per sector none of these splits is a result in its own right.
Changing only the search country on an identical question moved ChatGPT's sources almost completely. Overlap between the UK and US versions was 0.00 and 0.02 on the two questions, against ChatGPT's agreement with itself of 0.36 to 0.62 on those questions. Perplexity's UK and US lists overlapped by 0.64 and 0.43. With two pairs per engine this carries no statistical power, and we offer no Perplexity self-agreement figure to read it against, because its source list repeated exactly between runs.
What does a low overlap change about where you put effort?
It changes what you check: a result on one engine is no evidence for the other until you look.
Our AI Visibility Audit is a human review of your site plus a fixed panel of buyer prompts, tested at baseline and again at retest, and it records the sources each platform cites.
Get your AI Visibility audit →
The result in brief
Between-engine overlap in our study was a small fraction of what ChatGPT shared with its own same-day re-run, so the engines differ by far more than an engine differs from itself. Published figures for the pair run from about 6% to 13% because the studies count different things, and each engine needs its own measurement on your questions.
Frequently asked questions
Does a citation on Perplexity help me on ChatGPT?
We did not test it. Finding out would take your question set run on both engines before and after one change, with each engine's own repeat-run spread as the yardstick.
Is one run on each engine enough?
No. With a single run there is no way to tell an engine disagreeing with its rival from an engine disagreeing with itself. Two same-day runs give a first yardstick, and a retest days later shows how far the lists drift.
Should I report overlap by domain or by exact page?
Report both, labelled. Domain overlap tells you whether two engines draw on the same sites. Page overlap tells you whether they pick the same article on a shared site, and it comes out lower because many pages sit inside each domain.
Should questions that name my business be counted with the rest?
No. They test whether an engine can describe you once asked about you, which is recognition and sits apart from discovery.
Can I run this without paid software?
Yes. A spreadsheet handles the arithmetic. Collection is the slow part: paste each question into each app, or call an API, and stay on one route within a comparison, so that the engine is the only thing that changes.
Do your figures apply to my sector?
Only as a worked example. They come from a small question set run through an API on one day, so even a business in one of our three sectors needs its own run.
Sources
- [1] Qwairy, "Same Question, Different Web: AI Engines Barely Cite the Same Sources (June 2026)". https://www.qwairy.co/blog/ai-engines-citation-overlap-june-2026
- [2] Wellows, "89% of What ChatGPT Cites, Perplexity Never Touches". https://wellows.com/research/ai-citation-overlap-study/
- [3] Writesonic, "Do AI Engines Cite the Same Sources? We Studied 161,286 Prompts Across 4 Platforms". https://writesonic.com/blog/ai-citation-source-overlap-study
- [4] RankSurf, "ChatGPT vs Perplexity Citations: A 1,750-Answer Study". https://ranksurf.com/blog/what-chatgpt-and-perplexity-cite
- [5] Machine Relations Research, "Multi-Engine AI Citation Overlap: Among Sources Cited 10+ Times, 98.5% Reach Two or More Answer Engines". https://machinerelations.ai/research/multi-engine-ai-citation-overlap-data-2026
Leave a comment
Thoughts on this post? Leave a comment below. Comments are moderated before they appear, so yours will not show on the page straight away.
Your email is used only to contact you about your comment if needed; it is never published.
Comments
No comments yet. Be the first to leave one above.