ChatGPT, Perplexity, Gemini and Claude rarely cite the same page for the same buyer question
A study of 2,086 answers from the four engines to buyer-intent prompts found that 87.3% of the 17,483 pages cited for a prompt in a single run came from only one engine.
Summary
Strandline Research analyzed 2,086 answers that ChatGPT, Perplexity, Gemini and Claude gave to buyer-intent prompts for 12 brands between July 27 and September 9, 2026, and the 22,141 citations attached to them. For the same prompt in the same run, 87.3% of the 17,483 cited pages came from only one engine, and 0.1% were cited by all four. Perplexity attached 18.9 sources per answer and ChatGPT 3.0, and ChatGPT cited no source in 33.8% of its 530 answers. When the same 10 prompts were run daily on 43 dates, ChatGPT cited 23.9% of the same pages again the next day (n = 1,182) and Gemini 21.8% (n = 4,302), against 82.3% for Perplexity (n = 7,841). Listicle-style pages, such as best-of lists, alternatives pages and head-to-head comparisons, drew 37.5% of all 22,141 citations.
Findings
- For the same prompt in the same run, 87.3% of the 17,483 pages the four engines cited came from only one engine, and 26 pages (0.1%) came from all four (n = 474 prompt-runs).
- Perplexity attached 18.9 sources per answer and ChatGPT 3.0, a ratio of 6.4 (n = 540 and 530 answers), with Gemini at 11.6 and Claude at 8.7.
- ChatGPT cited no source in 33.8% of its 530 answers, and in 96.2% of its 106 answers to prompts that describe a problem.
- When the same 10 prompts were run daily on 43 dates, ChatGPT cited 23.9% of the same pages again the next day (n = 1,182) and Gemini 21.8% (n = 4,302), against 46.3% for Claude (n = 3,112) and 82.3% for Perplexity (n = 7,841).
- Listicle-style pages, such as best-of lists, alternatives pages and head-to-head comparisons, drew 37.5% of all 22,141 citations and 66.5% of the 6,481 citations for prompts asking for alternatives to a named competitor.
- Perplexity cited Reddit in 52.8% of its 540 answers and Claude in none of its 500, and Perplexity's answer text pointed to 8.1% of the 443 Reddit sources it listed, against 49.8% of all 10,226 sources it listed.
- The 10 most-cited domains drew 14.8% of the 22,141 citations, and 57.9% of the 3,193 cited domains were cited exactly once.
The four engines rarely cite the same page
Each run sends a project's prompts to all four engines at the same time. For every prompt-run, we pooled the pages the four engines cited and counted how many engines cited each page. Of 17,483 pages, 15,265 (87.3%) were cited by one engine only, 2,218 (12.7%) by two or more, and 26 (0.1%) by all four (n = 474 prompt-runs in which all four engines answered).
Overlap is higher at the domain level and still low. Of 15,026 domains, 2,855 (19.0%) were cited by two or more engines for the same prompt in the same run, and 55 (0.4%) by all four. The share is the same in both parts of the dataset: 19.0% of 11,702 domains in the daily-scan project and 19.0% of 3,324 domains in the 11 single-scan projects.
Strandline Research analysis of Cited scan data, July 27 to September 9, 2026. n = 474 prompt-runs in which all four engines answered. The URLs each engine cited are counted in the CSV.
ChatGPT overlaps least with the others. Of the 1,403 URLs it cited in these prompt-runs, 147 (10.5%) were also cited by another engine for the same prompt, against 1,348 of 4,092 (32.9%) for Claude. Taken one pair at a time, ChatGPT shared 1.3% of URLs with Claude (69 of 5,426 cited by either), 0.9% with Gemini (65 of 6,860) and 1.1% with Perplexity (111 of 10,371). The closest pair was Claude and Perplexity, at 10.1% (1,213 of 11,958).
| Engine pair | URLs cited by either | Cited by both | Share |
|---|---|---|---|
| Claude and Perplexity | 11,958 | 1,213 | 10.1% |
| Gemini and Perplexity | 13,467 | 1,134 | 8.4% |
| Claude and Gemini | 9,172 | 442 | 4.8% |
| ChatGPT and Claude | 5,426 | 69 | 1.3% |
| ChatGPT and Perplexity | 10,371 | 111 | 1.1% |
| ChatGPT and Gemini | 6,860 | 65 | 0.9% |
A page that earns a citation in one engine's answer is usually missing from the other three engines' answers to the same question on the same day. Visibility in one engine says little about the others.
Perplexity attaches 6.4 times as many sources per answer as ChatGPT
Strandline Research analysis of Cited scan data, July 27 to September 9, 2026. n = 2,086 answers: ChatGPT 530, Claude 500, Gemini 516, Perplexity 540.
| Engine | Answers | Citations | Mean per answer | Median per answer | Answers with no citation |
|---|---|---|---|---|---|
| Perplexity | 540 | 10,226 | 18.9 | 20.0 | 0.0% |
| Gemini | 516 | 6,005 | 11.6 | 10.0 | 10.3% |
| Claude | 500 | 4,331 | 8.7 | 9.0 | 1.4% |
| ChatGPT | 530 | 1,579 | 3.0 | 4.0 | 33.8% |
Perplexity's source list never ran past 20 URLs, the maximum across its 540 answers, and the median answer carried exactly 20. Gemini's longest list had 45 URLs. When ChatGPT did cite, it cited 4.5 URLs per answer.
The order holds in both parts of the dataset. In the 11 single-scan projects Perplexity averaged 18.1 citations per answer and ChatGPT 2.9 (n = 110 answers each). In the daily-scan project the figures were 19.2 (n = 430) and 3.0 (n = 420).
Part of the gap comes from how each engine reports sources. Perplexity returns a list of sources with every answer, and its answer text points to about half of them (see the Reddit section below). ChatGPT attaches a citation to the passages it wrote from a search result, and it can answer without searching at all.
ChatGPT answered most problem questions without a source
ChatGPT cited nothing in 179 of its 530 answers (33.8%). The rate depends on the kind of prompt. For prompts that describe a problem, 102 of 106 ChatGPT answers (96.2%) cited no source, against 65 of 265 (24.5%) for discovery prompts and 12 of 159 (7.5%) for prompts asking for alternatives to a competitor.
The pattern shows in both parts of the data: 84 of 84 problem-prompt answers (100.0%) in the daily-scan project and 18 of 22 (81.8%) in the single-scan projects. For the same problem prompts, the other engines nearly always cited something. Claude cited nothing in 7 of 100 answers (7.0%), Gemini in 11 of 103 (10.7%) and Perplexity in 0 of 108 (0.0%).
Three of the four engines replace most of their sources from one day to the next
One project's 10 prompts were run daily between July 27 and September 9, 2026, on 43 dates. For each engine and prompt, we compared the URLs cited on one day with the URLs the same engine cited for the same prompt on the next calendar day.
| Engine | Day pairs | URLs cited on day one | Cited again next day | URL retention | Domain retention |
|---|---|---|---|---|---|
| Perplexity | 410 | 7,841 | 6,453 | 82.3% | 84.1% |
| Claude | 356 | 3,112 | 1,441 | 46.3% | 50.6% |
| ChatGPT | 260 | 1,182 | 283 | 23.9% | 39.5% |
| Gemini | 323 | 4,302 | 936 | 21.8% | 27.3% |
ChatGPT's retention stays low when both days' answers carried citations: 25.5% across its 242 such day pairs. Its domain retention (39.5% of 1,068 domains) runs well above its URL retention, so it often came back to the same site with a different page.
- ChatGPT
- Claude
- Gemini
- Perplexity
Chart data
| days_later | ChatGPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|
| 1 | 23.9% | 46.3% | 21.8% | 82.3% |
| 2 | 23.2% | 45.5% | 20.8% | 78.2% |
| 3 | 21.3% | 43.7% | 20.7% | 75.6% |
| 4 | 21.3% | 43.6% | 20.7% | 74.0% |
| 5 | 20.3% | 44.3% | 19.0% | 72.2% |
| 6 | 20.6% | 44.3% | 19.4% | 70.8% |
| 7 | 17.6% | 43.1% | 17.4% | 69.8% |
| 8 | 17.0% | 42.6% | 19.0% | 68.6% |
| 9 | 16.0% | 41.1% | 17.8% | 67.8% |
| 10 | 15.7% | 42.1% | 15.8% | 67.1% |
| 11 | 12.0% | 40.5% | 15.4% | 66.2% |
| 12 | 13.6% | 41.5% | 16.3% | 65.5% |
| 13 | 12.0% | 41.8% | 15.5% | 64.1% |
| 14 | 11.6% | 40.3% | 14.4% | 64.0% |
| 15 | 13.4% | 39.3% | 14.3% | 63.2% |
| 16 | 10.3% | 39.9% | 13.7% | 62.2% |
| 17 | 9.9% | 38.9% | 15.0% | 61.0% |
| 18 | 9.3% | 38.6% | 15.8% | 60.9% |
| 19 | 9.9% | 38.6% | 15.8% | 61.4% |
| 20 | 9.7% | 37.3% | 15.7% | 61.0% |
| 21 | 10.1% | 39.6% | 16.5% | 59.1% |
| 22 | 8.9% | 38.6% | 17.5% | 58.3% |
| 23 | 9.6% | 38.4% | 18.2% | 57.5% |
| 24 | 11.3% | 38.3% | 17.2% | 57.5% |
| 25 | 7.4% | 36.7% | 16.9% | 57.1% |
| 26 | 8.7% | 38.4% | 17.8% | 57.1% |
| 27 | 9.4% | 38.0% | 17.2% | 56.6% |
| 28 | 8.7% | 36.2% | 18.3% | 56.1% |
Strandline Research analysis of Cited scan data. One project's 10 prompts, run daily on 43 dates between July 27 and September 9, 2026. The number of URLs behind each point is in the CSV.
The gap widens with time. Twenty-eight days apart, Perplexity still cited 56.1% of the same URLs (n = 3,178), Claude 36.2% (n = 1,103), Gemini 18.3% (n = 1,725) and ChatGPT 8.7% (n = 471).
Over the full period, Gemini cited 1,975 distinct URLs across the 10 prompts, 197.5 per prompt, while citing 11.7 per answer, and 65.3% of those URLs appeared on one day only. Perplexity cited 626 distinct URLs, 62.6 per prompt, and 19.0% of them appeared on one day only.
A single check of ChatGPT or Gemini shows one draw from a pool of pages that changes daily. These figures come from one B2B services brand and its 10 prompts. Other categories may turn over at different rates.
Listicle-style pages draw more than a third of citations
We sorted every citation into one source type with written rules, listed in the Method section. A URL counts as listicle-style when its path contains best or top followed by a separator or digit, alternative or alternatives, competitor or competitors, vs or versus, or compare or comparison, and its domain is outside the named platforms.
Strandline Research analysis of Cited scan data, July 27 to September 9, 2026. n = 22,141 citations. Classification rules are in the Method section.
| Source type | All engines | ChatGPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|---|
| Listicle-style pages | 37.5% | 39.6% | 47.7% | 29.6% | 37.4% |
| Review and directory sites | 3.8% | 9.6% | 3.1% | 3.5% | 3.5% |
| 2.4% | 1.5% | 0.0% | 1.0% | 4.3% | |
| YouTube | 1.0% | 0.0% | 0.0% | 0.5% | 1.8% |
| 0.9% | 0.6% | 0.0% | 0.1% | 1.8% | |
| News and business media | 0.7% | 0.4% | 1.4% | 0.7% | 0.4% |
| Medium | 0.1% | 0.0% | 0.1% | 0.3% | 0.1% |
| Own site | 0.1% | 0.5% | 0.1% | 0.1% | 0.1% |
| Wikipedia | 0.0% | 0.1% | 0.0% | 0.1% | 0.0% |
| All other pages | 53.5% | 47.6% | 47.5% | 64.1% | 50.6% |
| Citations (n) | 22,141 | 1,579 | 4,331 | 6,005 | 10,226 |
Listicle-style pages drew 8,298 of 22,141 citations (37.5%). Claude leaned on them most: they made up 47.7% of its 4,331 citations, and 461 of its 500 answers (92.2%) cited at least one. Gemini leaned on them least, at 29.6% of 6,005 citations.
The share follows the prompt. For prompts asking for alternatives to a named competitor, listicle-style pages drew 66.5% of 6,481 citations. For discovery prompts the share was 32.7% of 11,089, and for problem prompts 7.9% of 4,571, where 83.8% of citations went to other pages.
Review and directory sites such as G2, Clutch and Capterra drew 3.8% of all 22,141 citations and 9.6% of ChatGPT's 1,579. Named news and business media drew 0.7% of all citations. Wikipedia drew 9 of 22,141 citations, which rounds to 0.0%. The brands' own sites drew 24 citations (0.1%). The remaining 11,840 citations (53.5%) went to pages outside these rules, such as company sites, product pages, guides and blog posts whose URLs do not match the listicle pattern.
Perplexity lists Reddit often and seldom draws on it, and Claude never cites it
Reddit was the only domain cited for all 12 projects, and reddit.com was the second most-cited domain overall, with 526 citations. How often an answer cited it depended on the engine.
| Engine | Answers | Cited Reddit | Cited YouTube | Cited LinkedIn |
|---|---|---|---|---|
| Perplexity | 540 | 52.8% | 22.6% | 27.0% |
| Gemini | 516 | 10.9% | 5.4% | 1.2% |
| ChatGPT | 530 | 4.2% | 0.0% | 1.7% |
| Claude | 500 | 0.0% | 0.0% | 0.4% |
Claude cited Reddit and YouTube in none of its 500 answers. The data records only what each answer cited, so it cannot show whether Claude's searches returned Reddit pages that went unused.
Perplexity's answers carry numbered markers such as [3] that point into its source list, which shows which listed sources the text drew on. Across the 10,226 sources it listed, the text pointed to 5,095 (49.8%). Rank in the list mattered: the first-listed source was referenced in 459 of 540 answers (85.0%), the twentieth in 100 of 414 (24.2%).
Reddit sources sat lower in the list, with a median rank of 14.0 against 10.0 for all sources, and the text pointed to 36 of 443 (8.1%). Rank does not explain the gap. Among sources listed in the top five, the text pointed to 20 of 99 Reddit sources (20.2%) and 1,862 of 2,600 sources from other domains (71.6%). YouTube (29 of 183, 15.8%) and LinkedIn (49 of 180, 27.2%) also sat below the average. A Reddit thread in Perplexity's source list was usually one the answer text did not point to.
Citations spread across a long tail of domains
The 3,193 cited domains form a long tail. The 10 most-cited domains drew 3,283 of 22,141 citations (14.8%), and the top 1% of domains, 32 of them, drew 6,906 (31.2%). Half of all citations went to 89 domains, and 1,849 domains (57.9%) were cited exactly once. ChatGPT was the most concentrated engine: its 10 most-cited domains drew 571 of its 1,579 citations (36.2%).
| Rank | Domain | Citations | Share of all citations | Engines citing | Projects citing |
|---|---|---|---|---|---|
| 1 | pipeline.zoominfo.com | 545 | 2.5% | 2 | 3 |
| 2 | reddit.com | 526 | 2.4% | 3 | 12 |
| 3 | salesforce.com | 336 | 1.5% | 4 | 5 |
| 4 | g2.com | 304 | 1.4% | 4 | 9 |
| 5 | belkins.io | 286 | 1.3% | 4 | 1 |
| 6 | clicksgeek.com | 282 | 1.3% | 4 | 3 |
| 7 | saleshandy.com | 282 | 1.3% | 4 | 2 |
| 8 | apollo.io | 245 | 1.1% | 3 | 2 |
| 9 | cience.com | 242 | 1.1% | 3 | 1 |
| 10 | cognism.com | 235 | 1.1% | 4 | 3 |
| 11 | leadfeeder.com | 233 | 1.1% | 3 | 4 |
| 12 | salesbread.com | 228 | 1.0% | 3 | 1 |
| 13 | qualified.com | 222 | 1.0% | 4 | 1 |
| 14 | clutch.co | 214 | 1.0% | 4 | 3 |
| 15 | youtube.com | 211 | 1.0% | 2 | 11 |
The top of this list reflects the dataset's makeup. The daily-scan project produced 17,641 of the 22,141 citations (79.7%), so B2B sales software and lead-generation services fill most of the top 50. Counting projects gives a different top: reddit.com was cited for all 12 projects, youtube.com for 11, and g2.com, linkedin.com and forbes.com for 9. Leaving out citations of the brands' own sites, 2,741 domains were cited for one project only.
Method
Source. The data comes from Cited, an AI-visibility tool that sends a brand's buyer prompts to four AI engines through their APIs and stores every answer and every cited URL. This report uses every run stored between July 27 and September 9, 2026 (US Eastern dates).
Sample. 54 runs for 12 brand projects, each with 10 prompts (120 prompts in all). One project was scanned daily, 43 runs on 43 dates. The other 11 were scanned once. The projects are B2B software products, B2B services, consumer health products and a personal website. The engines received 2,130 requests and returned 2,086 answers. The 44 failed requests are excluded. The answers carried 22,141 citations of 6,342 distinct URLs on 3,193 domains.
Prompts. Cited writes each project's prompts with a language model from the brand's website and competitor list, with an instruction to leave the brand's name out. Every project has 5 discovery prompts, which ask for the best options in a category, 3 competitor-alternatives prompts, which ask for alternatives to or comparisons with a named competitor, and 2 problem prompts, which describe a problem to solve. That makes 60, 36 and 24 prompts across the 12 projects. The brands, their prompts and the answer text are not published.
Engines.
| Engine | Model | Search | Sources recorded |
|---|---|---|---|
| ChatGPT | OpenAI chat-latest, Responses API | Web search tool, used at the model's discretion | URL citations attached to the answer text |
| Perplexity | sonar | Built in | The citations list returned with the answer |
| Gemini | gemini-2.5-flash | Google Search grounding | Grounding sources, with Google redirect links resolved to their destination |
| Claude | claude-sonnet-5 | Web search tool, up to 4 searches per request | Citations attached to the answer text, or the search results when the text carried none |
Any URL written out in an answer's text was added to its list. Each engine answered each prompt once per run. Claude received a system prompt asking it to search the web, name specific products or services and cite the pages it used. The other three engines received the prompt alone. Claude answers begin on July 29, 2026. When an engine failed three times in a row within a run, the run stopped calling it, which is why Claude (512) and Gemini (538) have fewer than 540 requests.
Definitions.
- A citation is one URL attached to one answer. URLs are stored in canonical form, with tracking parameters, fragments and trailing slashes removed, and a URL that appears twice in one answer counts once. A page is a distinct URL.
- A domain is the URL's host name without "www". Subdomains count as separate domains in the domain tables. The source-type rules match a domain together with its subdomains.
- A prompt-run is one prompt in one run. The overlap analysis uses the 474 of 540 prompt-runs in which all four engines returned an answer.
- Retention is the share of URLs an engine cited for a prompt on one day that the same engine cited again for the same prompt on a later day. It is pooled across prompts and days. An answer with no citations on the later day counts as zero retained. The day-over-day figures use consecutive calendar days only.
- The daily-scan project and the single-scan projects are the two parts of the dataset used to check whether a result depends on the project mix.
- For Perplexity, a listed source counts as referenced when the answer text contains its numbered marker, such as [3] for the third source. Answer text was read only for this count.
Source-type rules. Each citation gets the first type whose rule matches.
- Own site: the URL is on the project's own domain or a subdomain of it.
- Reddit (reddit.com), YouTube (youtube.com, youtu.be), LinkedIn (linkedin.com), Wikipedia (wikipedia.org) and Medium (medium.com), each with subdomains.
- Review and directory sites: g2.com, capterra.com, clutch.co, trustpilot.com, getapp.com, softwareadvice.com, trustradius.com, gartner.com, sourceforge.net, slashdot.org, alternativeto.net, softwareworld.co, saasworthy.com, designrush.com, goodfirms.co, themanifest.com, sortlist.com, upcity.com, featuredcustomers.com, comparably.com, producthunt.com, bbb.org, yelp.com, glassdoor.com, sitejabber.com and consumeraffairs.com.
- News and business media: forbes.com, entrepreneur.com, inc.com, fortune.com, fastcompany.com, hbr.org, businessinsider.com, cnbc.com, bloomberg.com, reuters.com, wsj.com, nytimes.com, washingtonpost.com, usatoday.com, cnn.com, nbcnews.com, cbsnews.com, abcnews.go.com, foxbusiness.com, apnews.com, axios.com, bbc.com, bbc.co.uk, theguardian.com, time.com, newsweek.com, npr.org, theatlantic.com, marketwatch.com, yahoo.com, techcrunch.com, theverge.com, wired.com, venturebeat.com, zdnet.com, techradar.com, pcmag.com and cnet.com.
- Listicle-style pages: the lower-cased URL path matches
(^|[/_.-])(best|top)(\d|[-_/.](?!practice))|alternatives?([-_/.]|$)|competitors?([-_/.]|$)|[-_/](vs|versus)[-_/.]|compar(e|ison). The lookahead keeps "best-practices" pages out. - All other pages.
Exclusions. The 44 failed requests are excluded from every figure. Citations of a brand's own site (24, or 0.1% of 22,141) stay in all totals and shares and are left out of the domain lists, because a brand's own domain would identify it.
Limits.
- Twelve brands in a handful of categories. One project produced 79.7% of citations, so pooled figures lean toward its category. Where a result could depend on that mix, the report shows both parts of the dataset.
- The day-over-day results come from one project's 10 prompts.
- The engines were called through their APIs. Consumer apps can give different answers, with signed-in history, location, and their own models and settings.
- Answers vary from one request to the next. Each engine answered each prompt once per run, so any single answer is one draw.
- The engines report sources differently. Perplexity returned at most 20 sources per answer here, ChatGPT attaches citations only to text it wrote from a search, Gemini redirect links that failed to resolve were dropped, and Claude's list falls back to its search results when its text carries no citations. Counts per answer partly reflect these differences.
chat-latestis an alias, and the model behind it may have changed during the study.- The URL rules are heuristics. A comparison page whose URL lacks the listed words counts under All other pages, and a few matching URLs are guides.
Data
Every table and chart in this report comes from the CSV files listed below, and each file's description says what one row holds. Counts are plain integers and percentages are plain numbers rounded to one decimal. Engine names match the tables above. Projects appear only under anonymous labels (P01 to P12), and citations of the brands' own sites are left out of the two domain lists. The SQL that produced every file, including the source-type rules, is in queries.sql.
- Dataset totals One row: runs, projects, prompts, requests, answers, failed requests, citations, distinct URLs and domains, own-site citations and their share, and the first and last answer dates (US Eastern).
- Projects One row per project under an anonymous label (P01 to P12): scan schedule, runs, prompts, answers, failed requests, citations and share of all citations.
- Prompt types One row per prompt type (discovery, competitor alternatives, problem): prompts, prompts per project, answers and citations.
- Answers and citations by engine One row per engine: model id, requests, failed requests, answers, citations, mean, median and maximum citations per answer, answers with no citation and their share, citations per answer that cites at least one source, distinct domains and URLs, citations per answer relative to ChatGPT, and first and last answer dates.
- Engine summary by part of the dataset The engine summary for the daily-scan project and for the 11 single-scan projects: answers, citations, mean and median citations per answer, and answers with no citation.
- Answers with no citation, by prompt type One row per part of the dataset (all projects, daily-scan project, single-scan projects), engine and prompt type: answers, answers with no citation and their share.
- Cross-engine overlap For prompt-runs in which all four engines answered: distinct URLs or domains cited, and how many were cited by one engine, by two or more and by all four, for all projects and each part of the dataset.
- Overlap by engine For each engine: URLs and domains it cited in prompt-runs where all four engines answered, and how many another engine also cited for the same prompt in the same run.
- Overlap by engine pair For each pair of engines: URLs and domains cited by either and by both for the same prompt in the same run, summed over prompt-runs.
- Day-over-day retention Daily-scan project only. One row per engine: days answered, answers, URLs per answer, consecutive-day pairs, URLs cited on day one and still cited the next day, URL retention overall and when both days carried citations, domain retention, distinct URLs over the full period and the share cited on one day only.
- Retention by days elapsed Daily-scan project only. One row per gap of 1 to 28 days: the share of cited URLs the same engine cited again for the same prompt after that many days, per engine, with the number of URLs behind each share.
- Citations by source type One row per source type: citations and share for all engines, and citations and share for each engine.
- Source types by prompt type One row per prompt type and source type: citations and share of that prompt type's citations.
- Answers citing each source type One row per engine and source type: answers, answers citing at least one source of that type, and their share.
- Top 50 cited domains The 50 most-cited domains, citations of brands' own sites excluded: rank, citations, share of all 22,141 citations, number of engines and of projects citing the domain, domain group, and the share of the domain's citations whose URL matches the listicle rule.
- Domains by number of projects citing them One row per number of projects (1 to 12): domains cited for that many projects and their citations, with the domains named for 8 or more projects. Own-site citations excluded.
- Concentration of citations For all engines and each engine: citations, distinct domains, citations to the top 10 domains and to the top 1% of domains with their shares, the number of domains that account for half of citations, and domains cited once.
- Perplexity references by list rank One row per position in Perplexity's source list: sources listed, sources the answer text points to with a numbered marker, and their share.
- Perplexity references by source type One row per source type plus all sources: sources listed by Perplexity, sources the answer text points to, their share and the median list rank.
- Perplexity references in the top five Sources listed in ranks 1 to 5 of Perplexity's source list, Reddit against all other domains: sources listed, referenced in the text, and share.
How to cite: Strandline Research. "ChatGPT, Perplexity, Gemini and Claude rarely cite the same page for the same buyer question." 23 September 2026. https://strandlineresearch.com/reports/what-ai-engines-cite