The AI Citation Index

The domains AI answers actually cite

Ranked by breadth — how many separate accounts, in unrelated industries, saw the domain cited in their own scan. Raw citation volume is the wrong sort key: a domain quoted 1,400 times inside three scans of one niche tells you about the niche, not about the engines. By that measure the most universal source in AI answers is reddit.com, which showed up in scans belonging to 71 of 74 accounts.

Every engine has a different diet

Share of each engine’s citations by the kind of page it pointed at. The spread is the story: these are not six front-ends over one index, they are six different retrieval philosophies. Perplexity sends 3.5% of its citations to forums and community threads; Claude sends 0.3%.

Source type	ChatGPT	Claude	Gemini	Perplexity	AI Overviews	Grok	All engines
Community & forums Threads real people wrote — Reddit, Quora, Stack Exchange, Hacker News.	3.0%	0.3%	0.8%	3.5%	2.8%	1.5%	2.5%
Video Video platforms, where the citation points at a watch page or transcript.	0.2%	0.1%	0.7%	2.8%	6.7%	1.0%	2.1%
Social networks Profile and post URLs on the big social graphs.	0.4%	1.1%	0.2%	4.7%	5.0%	3.0%	3.2%
Reviews & directories Software review marketplaces, agency directories and ratings sites.	3.1%	4.0%	1.0%	2.8%	1.1%	1.8%	2.6%
Encyclopedias & reference Wikipedia and the other general-reference encyclopedias.	2.4%	0.7%	0.3%	0.4%	0.3%	0.7%	0.6%
News & media Publishers and newswires, from national press to trade titles.	0.5%	0.8%	1.0%	0.8%	0.6%	1.5%	0.8%
Academic & government Journals, preprint servers, .edu, .gov and intergovernmental bodies.	4.8%	3.5%	2.4%	2.8%	1.1%	3.4%	3.0%
Statistics aggregators Sites whose whole product is scraped statistics round-ups.	0.3%	0.3%	0.1%	0.9%	0.1%	0.1%	0.5%
Vendor & independent sites Residual — everything the lists above didn’t recognise. Not shaded: it would flatten every other row.	85.2%	89.2%	93.6%	81.3%	82.3%	86.9%	84.7%

Ranking in one engine barely predicts the next

Give two engines the identical prompt in the identical scan, and compare the sets of domains they cite. The number below is the mean Jaccard overlap — shared domains as a share of all domains either engine used. Gemini and ChatGPT agree on 4.5% of their sources. The best-agreeing pair on the board, Grok and ChatGPT, manages 15.1%. There is no such thing as “ranking in AI”.

Overlap	ChatGPT	Claude	Gemini	Perplexity	AI Overviews	Grok
ChatGPT	—	6.1%	4.5%	6.4%	9.6%	15.1%
Claude	6.1%	—	8.7%	13.3%	10.0%	10.5%
Gemini	4.5%	8.7%	—	9.1%	9.6%	11.7%
Perplexity	6.4%	13.3%	9.1%	—	13.1%	12.9%
AI Overviews	9.6%	10.0%	9.6%	13.1%	—	12.4%
Grok	15.1%	10.5%	11.7%	12.9%	12.4%	—

How often each engine cites anything at all

Before overlap can matter, an engine has to show its sources. AI Overviews attaches sources to 85% of its answers; ChatGPT to 30%.

Engine	Answers with sources	Answers	Citations	Sources per cited answer
AI Overviews	85%	3,420	27,599	9.5
Perplexity	67%	17,314	176,212	15.3
Grok	45%	17,273	36,976	4.7
Claude	38%	9,530	46,649	12.8
Gemini	37%	17,528	38,094	5.8
ChatGPT	30%	17,502	31,390	6.1

AI does not cite new content

For every page we saw cited and could fetch, we read the publish date out of its own markup and compared it to the first time we recorded an engine citing it. The median page was 196 days old. Only 9% were under a month old; 69% were under a year. Publish-and-wait is not a strategy, but neither is judging a page after three weeks.

What a cited page is made of

We fetch the pages engines cite and pull their structure apart. This is the composite of 4,627 of them — not what SEO advice says a citable page should look like, but what the pages winning citations right now actually contain.

Format of the cited page

How this is measured, and what it can’t tell you

Questions about the data

Which sources do AI answer engines cite most?

Across 356,920 citations collected from six engines between April 2026 – September 2026, the domain seen by the most separate accounts is reddit.com — cited in scans belonging to 71 of 74 accounts. Ranking by breadth rather than raw volume matters: a domain cited thousands of times inside one niche is a property of that niche, not of the engines.

Does ranking in ChatGPT mean you rank in Gemini?

No. On the same prompt in the same scan, Gemini and ChatGPT share only 4.5% of their cited domains — the weakest agreement of any pair we measure. The strongest pair, Grok and ChatGPT, still only reaches 15.1%. Answer-engine visibility has to be measured per engine.

How long after publishing does AI start citing a page?

Longer than most content calendars assume. Of 2,052 cited pages with a readable publish date, the median was 196 days old when we first recorded an engine citing it, and only 9% were under a month old. That figure is an upper bound — we learn a page is cited when a scan runs, not the moment it happens.

What does a page that AI cites look like?

The median cited page runs 1,683 words, carries 15 discrete statistics and 10 self-contained quotable paragraphs, and 62% of them include an FAQ block. Guides and listicles dominate the format mix.

Can I use this data?

Yes. Every table on this page has a CSV download and the data is free to republish with attribution and a link to rankmonster.ai/sources. If you want a cut we do not publish — a specific category, a longer window, a different slice — email [email protected] and we will usually run it.

How is this measured?

We run fixed prompt sets through six answer engines and record every source URL each engine returns with its answer. This edition covers 82,567 engine answers across 7,920 distinct prompts and 578 scans, from 74 accounts. Citations are counted per engine response; domains keep their subdomain.