What AI engines cite: the first eight weeks of the AI Source Index

1

The engines rarely cite the same sources for the same question.

In the 283 of 320 week-and-question cells that every engine answered, 6,264 of the 7,936 distinct domains cited — 78.9% (95% interval 77.2–80.6%) — were cited by just one engine. 7 were cited by every engine.

Pair by pair, two engines shared 3.6% to 11.4% of the domains either of them cited for the same question in the same week. The same engine, asked the same question one week later, shared 30.8% (29.6–32.0%) with itself.

Distinct domains per week-and-question cell, by how many engines cited them

• 1 engine6,264
• 2 engines1,136
• 3 engines379
• 4 engines123
• 5 engines27
• 6 engines7

Overlap between two engines: shared domains ÷ domains either cited, same question, same week

• Same engine, one week apart30.8% · 4,622/15,001
• AI Overviews × Google AI Mode11.4% · 215/1,878
• Gemini × AI Overviews10.7% · 431/4,016
• Perplexity × Claude10.5% · 321/3,069
• Gemini × Claude9.0% · 422/4,674
• Perplexity × Gemini7.9% · 319/4,025
• Perplexity × AI Overviews7.9% · 201/2,540
• Gemini × Google AI Mode7.9% · 271/3,425
• Claude × AI Overviews7.5% · 243/3,250
• Perplexity × Google AI Mode6.5% · 121/1,869
• Claude × Google AI Mode6.2% · 160/2,582
• Perplexity × ChatGPT4.7% · 141/3,012
• ChatGPT × Gemini4.1% · 191/4,668
• ChatGPT × Claude3.9% · 146/3,759
• ChatGPT × Google AI Mode3.8% · 92/2,413
• ChatGPT × AI Overviews3.6% · 112/3,144

Why it matters: A result in one engine is not evidence about another. Measure each engine your buyers use on its own.

2

Reddit and YouTube are the most-cited domains, but only some engines cite them.

Gemini cited reddit.com in 193 of its 320 answers (60.3%; 95% interval 51.6–69.4%). ChatGPT and Claude did not cite it once in 8 weeks. Wikipedia appeared in 19 of all 1,883 answers.

Answers citing each platform at least once, by engine (all eight weeks)

A platform counts when an answer cites its domain or a subdomain of it (m.youtube.com, en.wikipedia.org). Each cell: answers citing it at least once, out of the answers that engine gave. No interval is printed for a zero: resampling questions that never cited a platform can only produce zero again, so the bootstrap cannot say how rare it is.

Read with care: ChatGPT here is OpenAI’s gpt-4o-mini through the API with web search, and Claude is claude-haiku-4-5-20251001 through Anthropic’s API with its search tool. Our instrument log records Reddit disappearing from the ChatGPT lane on 31 August 2026, the day that lane moved to gpt-4o-mini. These zeros describe these configurations on these questions, not the consumer apps.

Why it matters: “Get discussed on Reddit” reaches Google’s two surfaces and Gemini on these questions and little else. YouTube travels further: ChatGPT cited it in 76 of 320 answers. Check which engines your buyers use before planning around one platform.

3

Beyond those two, citations spread over a very long tail.

The answers cited 2,669 different domains. The ten most cited took 14.1% of all 11,390 domain citations (1,611), and youtube.com and reddit.com 7.8% on their own. It takes the top 259 domains to cover half of all citations, and 1,181 domains — 44.2% of them — were cited exactly once in eight weeks.

The 10 most-cited domains, of 2,669

• 1. youtube.com social499
• 2. reddit.com community386
• 3. forbes.com media145
• 4. techradar.com media125
• 5. nerdwallet.com media89
• 6. zapier.com other84
• 7. tomsguide.com media78
• 8. cnbc.com media72
• 9. rtings.com media67
• 10. northflank.com other66

Counts are answers citing the domain at least once, across all engines and all eight weeks. Subdomains count separately (m.youtube.com is not youtube.com), as in the weekly index. Type is the index’s source taxonomy, applied for reading only.

Why it matters: No short list of sites accounts for most AI citations on these questions: half of them went to domains outside the top 259.

4

About half of what an engine cites for a question changes from one week to the next.

Of the domains an engine cited for a question one week, 46.8% were cited again for the same question the next week (4,622 of 9,868; 95% interval 45.5–48.2%) — from 37.4% on Google AI Mode to 53.2% on Claude.

Over all eight weeks, for the 203 of 240 question-and-engine pairs answered every week, 48.1% of the 4,112 question-engine-domain combinations ever cited appeared in only one week, and 2.0% in all eight.

Cited again the next week, by week pair (all engines)

• W36 → W3755.2% · 782/1,416
• W37 → W38 (instrument changed)46.0% · 580/1,262
• W38 → W3952.9% · 645/1,220
• W39 → W40 (instrument changed)44.6% · 703/1,575
• W40 → W4142.4% · 625/1,474
• W41 → W4243.9% · 659/1,501
• W42 → W4344.2% · 628/1,420

Cited again the next week, by engine (all week pairs)

• Perplexity50.5% · 669/1,324
• ChatGPT49.0% · 889/1,815
• Gemini46.7% · 1,338/2,865
• Claude53.2% · 889/1,671
• AI Overviews38.5% · 584/1,516
• Google AI Mode37.4% · 253/677

Counting only week pairs measured with the same instrument fingerprint: 47.5% (3,339 of 7,031). The week pairs this window shares with our 7 October persistence analysis reproduce it exactly. The earlier analysis: how long AI citations last

Note added 26 October 2026: one of those pairs, W40→W41, shares the instrument fingerprint but spans a change of the collector from version 2 to 3, which our weekly pages treat as not comparable. Without it the rate is 48.8% (2,714 of 5,557; 95% interval 47.0–50.7%). The finding does not change: about half of the cited sources are kept from one week to the next. The archived files are unchanged; the per-pair counts above add up to this figure.

Why it matters: One week’s answer is one draw. Judge a change in your citations over several weeks, never from a single check.

5

YouTube is cited far more on head-to-head “X vs Y” questions than on any other kind.

179 of 376 answers to the 8 “X vs Y” questions cited YouTube (47.6%; 95% interval 42.4–52.9%), against 19.6–23.3% on the other three kinds of question.

Answers citing YouTube and Reddit, by question shape (all engines)

Read with care: Each of these shapes has only 8 questions, so the intervals are wide. The “X vs Y” interval does not overlap any other shape’s; the exact size of the gap is not known.

Why it matters: When buyers compare you head-to-head with a rival, video comparisons are part of what the engines cite on these questions. That is a pattern in this panel, not a tested lever.

6

Engines differ several-fold in how many sources one answer cites.

Gemini cited 10.33 domains per answer on average (95% interval 9.61–11.05); Google AI Mode cited 2.37 (1.84–2.93), and 107 of its 320 answers (33.4%) cited no countable source at all.

Sources per answer and answers citing nothing, by engine

Read with care: Part of AI Mode’s low count is our counting rule, not the engine: Google’s own search and shopping links are not counted as sources, so an answer that linked only to them counts as citing nothing.

Why it matters: “Cited in X% of answers” means different things on different engines: one that lists ten sources has more slots than one that lists two. Compare engines with that in mind, not by raw rates alone.

Method

40 fixed English buyer questions — 16 “best …”, 8 “alternatives to …”, 8 “X vs Y” and 8 “how to choose …” — with the same wording every week and none about Promvia. Each week each question was put once to each answer engine (the first run on Thursday 3 September, every later one on a Monday from 7 September): Perplexity, ChatGPT, Gemini and Claude through their APIs with web search, Google AI Overviews and AI Mode read from Google’s results page through a SERP provider. One answer is one (week, question, engine) check; a domain counts once per answer; a failed check is coverage, never a citation. All 8 weeks were fetched from the index’s public rows endpoint on the same day and are archived unchanged.

Every interval is a 95% percentile bootstrap that resamples the 40 questions with replacement (2,000 draws). The same questions recur every week and across engines, so the answers are clustered; an interval that treated each answer as independent would be too narrow.

Full method: how we measureThe weekly AI Source Index

What could fool this measurement, and what we did about it

• False zeros. The questions are broad category questions. A domain absent here was not cited for these questions, which is not the same as never cited; nothing in this report says a site is invisible to AI.
• Echoes. We count cited links, not names in the answer text, so an engine repeating a brand name from the question cannot count as a citation. Answers do link to the sites of brands the question names (vercel.com for “Vercel vs Netlify”), and those count like any other citation. Our automatic flag for them also matched look-alike domains, so we publish no figure for them.
• Model changes. The rows record the model from 2026-W38: ChatGPT gpt-4o-mini, Gemini gemini-3.1-flash-lite and Claude claude-haiku-4-5-20251001, unchanged to the end of the window. The two earlier weeks predate the field; our instrument log puts the ChatGPT lane on its model from 31 August 2026, before the window opens. Perplexity is asked through a preset and Google’s two surfaces through their results pages, so no model is recorded for them, and Google can change what powers them without notice. The ChatGPT lane moved to gpt-5-mini on 21 October 2026, after this window, so later weeks are a different instrument.
• Failed checks. 37 of 1,920 checks failed (Claude 29 in 2026-W38, other, AI Overviews 1 in 2026-W41, provider-error, AI Overviews 2 in 2026-W42, no-answer-parsed, AI Overviews 1 in 2026-W42, provider-error and Perplexity 4 in 2026-W43, rate-limited). They stay in the rows with their failure class and leave the denominator; a failed check never counts as “not cited”.
• Instrument changes. The panel’s fingerprint changed at 2026-W38, when provider settings began to be frozen into it, and at 2026-W40, when the Bing stand-in for Copilot left the panel. Week pairs across those dates are included in the pooled figures, and the same-instrument figure is given beside them.
• Microsoft Copilot is not in this report. Early in the window the panel carried a Bing organic-results stand-in for it, which failed every check it was given and was then removed. Its rows are left out of every count in this report, including the 1,920 checks and the failures, because it was never an answer engine in scope. That is unlike a failed check in scope, which stays in the rows. There is no Copilot measurement to report, and none is implied. A failure class of “other” means the provider’s error matched none of the named classes. The index stores the class only, never the provider’s message.

Limitations

• One sample per engine, question and week. AI answers vary from run to run; a single week’s citation is one draw, which is why this report pools weeks and prints intervals.
• A fixed set of English buyer questions in a handful of consumer and business categories, with Google’s surfaces read for a US location. They are not a sample of all questions, languages or markets, and the platform shares above belong to these questions.
• API and results-page surfaces, not the consumer apps. Several lanes use small, fast models; an answer in a signed-in app, with memory and location, can differ.
• Weekly resolution over eight weeks. What happens inside a week, or over a longer stretch, is not observed here.
• Named platforms only. The index’s source taxonomy is an allow-list that leaves most cited domains unclassified, so this report does not publish shares by source type.
• Observation, not cause. Nothing here says why an engine cites a source, or that changing a page would change what it cites. Promvia does not cause citations or rankings, and every pattern on this page is a correlation in this panel.
• Who runs it. Promvia sells an AI-visibility product. The panel is fixed, unrelated to our customers, and contains no question about us.

Data and recount

Every number on this page is computed from the published rows, twice, by two independent programs — one in TypeScript, one in Python — and is printed only because both agree on every value. A test in our repository recomputes the page’s results from the rows and fails on any difference.

• rows.csv — one line per answer and cited domain (11,667 lines); an answer that cited nothing, or failed, keeps one line with an empty domain
• results.json — every number on this page
• recount.py — the recount script (Python 3, standard library only)
• The weekly rows as the index’s public API serves them
• Archived on Zenodo with a DOI, the permanent identifier to cite: 10.5281/zenodo.23651937
• The AI Source Index data repository on GitHub

Recount it yourself

1. Download rows.csv, results.json and recount.py into one folder.
2. python3 recount.py rows.csv --check results.json
3. It prints ALL MATCH when every published value comes back identical, and names any value that does not.

How to cite this report

Promvia (2026). What AI engines cite: the first eight weeks of the AI Source Index (September–October 2026). Promvia AI Source Index, weeks 2026-W36 to 2026-W43. https://promvia.app/ai-source-index/first-eight-weeks. Data: https://doi.org/10.5281/zenodo.23651937

The data (rows.csv, results.json) is licensed CC BY 4.0, like the AI Source Index itself: reuse it for anything, with attribution to “Promvia AI Source Index” and a link to this page. recount.py is MIT-licensed, like the index’s scripts.

Permalink: https://promvia.app/ai-source-index/first-eight-weeks