Where ChatGPT and Perplexity get their sources
Ask ChatGPT the same question twelve times and the twelve answers, between them, cite four times as many hosts as any one of them. Ask Perplexity twelve times and the answers cite nearly the same list. That is the first thing 122,018 citations from 12,408 answers show, and it changes how to read any single screenshot of an AI answer. The rest of this page is what the two engines cite, by kind of site, and where they part ways.
ChatGPT (gpt-5-mini) and Perplexity (sonar) both put company sites first, about half of all citations each. After that they diverge. Perplexity reaches for LinkedIn, information portals and list pages; ChatGPT reaches for Wikipedia, government and data sites. ChatGPT's source set is unstable across repeats, so one ChatGPT answer shows roughly a quarter of the hosts twelve answers draw on.
Twelve ChatGPT answers cite four times the hosts one answer shows
For each prompt-month we counted the distinct hosts cited across the 12 repeats and divided by the mean number of hosts in one answer. A ratio of 1 means every repeat cites the same hosts; a ratio of 4 means one answer shows a quarter of the set.
| Engine | Prompt-months | Distinct hosts across 12 repeats (mean) | Hosts per answer (mean) | Ratio, median [IQR] | Host slots in at least half the repeats | In every repeat |
|---|---|---|---|---|---|---|
| ChatGPT | 383 | 24.1 | 5.8 | 4.0 [3.27, 4.82] | 14.2% | 1.8% |
| Perplexity, first 10 | 517 | 10.2 | 8.16 | 1.12 [1.0, 1.35] | 80.7% | 63.7% |
The ratio is computed per prompt-month and the table shows its median, with the interquartile range in brackets, not a confidence interval. A host slot is one host within one prompt-month. Only prompt-months where the engine cited something are included, so for ChatGPT this covers the prompt-months where it searched.
Put the other way round: 14.2% of the hosts ChatGPT cites for a prompt show up in at least half of the 12 runs, and 1.8% in every run. On Perplexity it is 80.7% and 63.7%. The mean pairwise Jaccard overlap between two runs' host sets is 0.277 on ChatGPT and 0.892 on Perplexity. Collapsing hostnames to registered domains, so that two subdomains of one site count once, moves ChatGPT's median to 3.8 [3.20, 4.62] and leaves Perplexity at 1.125 [1.00, 1.35].
Five checks on the stability gap
| Check | ChatGPT median ratio | Perplexity median ratio |
|---|---|---|
| First 5 hosts of each answer only | 4.00 | 1.20 |
| First 3 hosts of each answer only | 4.33 | 1.00 |
| Same 383 prompt-months on both engines | 4.00 | 1.11 |
| September only | 4.14 | 1.28 |
| October only | 3.83 | 1.00 |
The gap is not an artifact of list length or sample. Perplexity's own ratio depends on how many sources the API returned: on September's full lists of 20 the median is 1.20 (mean pairwise Jaccard 0.86), on September's first 10 it is 1.28 (0.84), and in October, when the API returned 10, it is 1.00 (0.95).
This is a 12-repeat figure. With 6 repeats the ChatGPT ratio is about 2.8, so a tracker that runs fewer repeats will report a smaller gap and will not be wrong. If you are checking whether an AI engine cites you, one ChatGPT answer is a sample, not a verdict: One check against twelve works out how many runs it takes.
Perplexity cites LinkedIn; ChatGPT almost never does
A LinkedIn URL appears among the first 10 sources of 16.0% [12.7, 19.3] of Perplexity answers, 991 answers in all. On ChatGPT it is 0.7% [0.5, 1.0], 45 answers. The LinkedIn row of the reach table above reads 0.9% on ChatGPT because that type also counts LinkedIn's regional and business subdomains. The figures in this section are the linkedin.com host alone. By answers LinkedIn is Perplexity's second host (the first has 1,113), by citation rows also second (1,406 against 1,635), and on the mean of per-industry shares it is first, at 18.0%.
The prompt matters more than the engine's taste. On the 230 recommendation-style prompts (which firms, best agencies for, who should we hire) LinkedIn reaches 20.0% [15.7, 24.2] of Perplexity answers; on the 101 other prompts, 5.8% [1.5, 10.5]. On a 66-prompt energy panel written for no tracked brand, LinkedIn is Perplexity's top host at 27.5% [18.3, 37.9] of 792 answers, with the next host at 10.0%, and it appears on 24 of the 66 prompts. ChatGPT ranks LinkedIn 21st on the same panel, at 2.5%.
It is spread, not a few prompts: LinkedIn appears on 82 of 331 prompts (24.8%), and the five prompts that cite it most hold only 12.1% of its answers. Between the two fills its per-industry rate moved by a factor of 2.0 up in one industry and 3.3 down in another, so treat any one month's LinkedIn share as a range.
Position matters on Perplexity. In September, when the API returned 20 sources, LinkedIn reached 17.6% of answers on the first 10 sources and 27.8% on all 20, 58% more; by industry that rise runs from 0% to 78%. LinkedIn sits late in Perplexity's lists.
| LinkedIn URL kind | Share of 2,231 LinkedIn URLs |
|---|---|
| Company pages | 32.3% |
| Posts | 25.5% |
| Pulse articles | 24.6% |
| Personal profiles | 14.3% |
| Other | 3.4% |
Both engines, every list position.
ChatGPT cites Wikipedia; Perplexity does not
A wikipedia.org page appears in 12.6% [10.6, 15.2] of ChatGPT answers and 0.4% [0.0, 1.0] of Perplexity's first-10 lists: 24 answers from 2 prompts (42 answers from 7 prompts on all 20 September sources). The paired difference is [10.1, 14.7] points, and ChatGPT's reference reach is higher in every one of the 7 industries. The composition table below shows the same thing at the level of citations: reference sites are 3.6% of ChatGPT's answer-host pairs and 0.2% of Perplexity's.
List pages: both engines cite them, Perplexity more
A list page here is one whose title matches a pattern for best, top, ranked and similar words. A hand check of 40 matched titles read 37 of 40 as list pages (92.5%, Wilson 80% to 97%, counting the unsure as wrong), and 9 of 40 unmatched titles were list pages the pattern missed, with 8 more unsure, so the pattern is a lower bound and the finding is directional.
Of 61,982 Perplexity first-10 citations, 19.1% [16.3, 21.9] point at a list page; of 27,572 ChatGPT citations, 10.8% [8.6, 12.9]. The ratio is 1.76, and it ranges from 1.74 to 2.4 across the pattern variants we tried. By answers, 51.0% [45.5, 56.2] of Perplexity answers cite at least one list page against 22.2% [18.7, 25.6] of ChatGPT answers, but that comparison counts all 10 of Perplexity's sources against ChatGPT's 4.3 to 4.5, so read the per-URL figure. Perplexity's first 3 sources include a list page in 35.6% of answers and its first 5 in 41.7%; ChatGPT's first 5, in answers with 5 or more sources, 27.4%. Among ChatGPT answers that cite anything at all, 33.0% (1,377) cite a list page.
The whole mix
| Host type | ChatGPT, % of 23,137 answer-host pairs | Perplexity first 10, % of 47,818 |
|---|---|---|
| Company site | 49.5% | 49.7% |
| Government | 9.3% | 2.1% |
| Data and research | 5.7% | 3.6% |
| Medical and academic | 4.1% | 2.4% |
| Directory | 4.1% | 5.3% |
| Reference (Wikipedia and similar) | 3.6% | 0.2% |
| Information portal | 3.4% | 9.8% |
| Review and affiliate | 1.7% | 3.9% |
| News and trade | 1.5% | 4.3% |
| Podcast | 1.3% | 1.5% |
| 1.0% | 0.3% | |
| 0.2% | 2.1% | |
| Unclassified | 13.1% | 12.0% |
Composition: each row is the share of answer-host pairs that fall on that type of host. A host that appears in one answer is one pair. Unclassified hosts are the 12.5% of pairs neither hand nor rule could type.
Company sites hold 49.5% and 49.7%: the same on both engines. The industry moves that share by up to 15 points; the engine does not. Two industries on the same engine differ by 0.34 (ChatGPT) to 0.38 (Perplexity) in total variation distance on average, and within one industry the two engines differ by 0.22. Directional, but the direction is clear: the industry a question belongs to moves the source mix more than the engine answering it does.
Table view: reach by host type
| Host type | ChatGPT, % of answers | 95% interval | Perplexity first 10, % of answers | 95% interval |
|---|---|---|---|---|
| Company site | 50.4% | [46.1, 54.3] | 88.2% | [85.1, 90.8] |
| Unclassified | 27.8% | [25.2, 31.0] | 47.4% | [43.2, 52.2] |
| Directory | 10.7% | 25.2% | ||
| Information portal | 9.5% | 25.2% | ||
| News and trade | 5.0% | 24.0% | ||
| Review and affiliate | 5.2% | 16.3% | ||
| 0.9% | 16.0% | |||
| Data and research | 10.7% | 11.9% | ||
| Government | 18.5% | 11.1% | ||
| Medical and academic | 8.8% | 9.8% | ||
| Press wire | 0.6% | 4.9% | ||
| Podcast | 2.6% | 4.6% | ||
| Finance and investor | 0.2% | 3.9% | ||
| Nonprofit and association | 3.2% | 3.9% | ||
| Other social | 0.6% | 3.1% | ||
| Marketplace | 0.1% | 2.4% | ||
| 3.9% | 2.0% | |||
| Reference (Wikipedia and similar) | 13.2% | 1.2% | ||
| YouTube | 0.2% | 1.0% |
Intervals on the two largest rows are prompt-clustered bootstraps (400 draws); the other rows are shares of 6,204 answers.
Composition answers what fraction of citations; reach answers what fraction of answers. A type can be small in composition and large in reach when it appears once in many answers, which is the LinkedIn pattern on Perplexity.
How hosts were typed
66.2% of answer-host pairs were typed by hand, 21.3% by rule (the top-level domain, words in the host, or the page titles the engine returned), and 12.5% stayed untyped, so coverage is 86.9%. A hand spot check of 343 rule-typed hosts puts the expected mistyping at 2.7% of all pairs, with a range of 1.2% to 6.4%. Three things are off this page on purpose: government and Reddit shares, which did not hold up under review; brands named but not linked, confounded by a crawler block on part of the sample; and host concentration, which measures repeat stability under another name. The instrument and the sample are on the method page.
What to do with it
- Treat a ChatGPT screenshot as one draw from a wide set. Twelve runs is the floor before anyone says that ChatGPT cites a given source.
- On Perplexity, a company LinkedIn page is a citation surface, most of all on which-firms questions. 32.3% of the LinkedIn URLs cited are company pages, not personal profiles.
- List pages reach half of Perplexity answers. Being on the lists the engine already cites is the lever this study measured; writing a new list is not.
- Pick the engine before picking the reference. Wikipedia, government and data sites are in ChatGPT's pool and nearly absent from Perplexity's first 10.
- Have a prompt you want in the November panel? Suggest it; the panel is chosen for the questions these pages raise, not for any client.
Quoting a figure? Link to this page as the source: https://www.bestseopodcast.com/research/ai-citation-sources. Figures are dated 2026-10-25 and cover answers collected 12 to 26 September 2026 and 1 to 5 October 2026; the next refresh is after the November 2026 fill closes, about 15 November. The page is the citation: there is no download.
Who did this
Matthew Bertram runs EWR Digital, the SEO and AI-visibility agency whose monthly tracker produced this corpus, and he is the author of LLM Visibility: A Decision-Grade System for Winning AI-Mediated Discovery. The studies were built, re-derived from counts and reviewed before anything was written up; every figure on this page traces to a script in the project record. Questions about the method go to him through the contact page. More about Matthew Bertram.
Which of these sources cite you?
The audit reads what ChatGPT and Perplexity say about your company and where the gaps are.
Get a free AI visibility audit