Seven Websites Get Cited Whatever You Sell. The Rest Depend on Your Category.

What We Measured

4,100 questions, each put to ChatGPT, Perplexity and Gemini inside the same tracking run, for 12,300 responses between 2026-07-06 and 2026-08-22. Answers move between identical runs, so an engine checked on Tuesday against one checked on Friday measures the calendar as much as the engine. Asking all three inside one run removes that. Counts are presented at 10 runs per question per engine, the scale our CRM study uses; the shares, which are what the tables actually report, are measured and unscaled.

Citations are counted one row per cited link, and four filters run before anything is ranked. The window starts in July because the corpus before the late-June model cutover measures retired engines, a call our Reddit-by-engine post explains in full. Every run is then checked against its own billing records and kept only if it ran on the engine's current model, which keeps out two that were live days before this window opened: a deliberately non-searching ChatGPT model, and the Gemini 3 preview we reverted.

Every answer must also have come back with at least one source. That is the only per-answer evidence that a live search happened: the search flag on our own billing rows is set for ChatGPT but not reliably for the other two, because Perplexity's per-request search fee went unmodeled until mid-August and Gemini's grounding fee is not modelled at all. On the sources test Gemini loses about a fifth of its answers and the other two lose almost none. Then every citation to the brand's own website, its subdomains and its alias domains is removed, so these are the websites engines cite about a company rather than the company itself. Own-domain pages are only 2.9% of AI citations in our published study, and they are the one kind of page a brand already controls.

Shares are then averaged company by company. We compute a website's share of each company's citations, then average across companies, so a company with many prompts cannot tip the list. Ranked by raw volume instead, the top of every list fills with exactly those sites: a medical reference for the one clinic in the sample, a startup directory for the one startup, a regional finance site for the one broker. A website is called general only if it was cited for at least half the companies on at least one engine, and websites below that bar are named nowhere in this post.

Claude and Grok are absent because they are Agency-tier engines with too few brands in the window to match, and Grok returns no citations since xAI retired Live Search. Google AI Overviews was checked and left out for the same reason as in the Reddit post: too few brands answer for it to publish a share.

What this sample is and is not. The companies behind these questions are mostly software and business services, with a visible tilt toward brands based in India and only a handful of consumer categories. A consumer-heavy corpus would rank retailers and video higher, which is what the public trackers show. So the general set below is the part of the picture that held across every category we track, and the type rankings are the part most likely to travel to yours.

The Seven Websites Every Category Gets Cited From

These are the only websites that cleared the bar on at least one engine. Each cell gives the site's average share of a company's citations on that engine.

Two rows carry almost all of the divergence. Reddit runs from 17.87% of a company's citations on ChatGPT to 9.07% on Perplexity and 3.43% on Gemini, a split our Reddit-by-engine post takes apart on its own. Wikipedia is the other: worth 1.95% on Perplexity and 0.27% on Gemini, which is the difference between a reference entry being a priority and a rounding error.

The rest is steadier than the public league tables suggest. YouTube lands between 1.93% and 3.77%, TechRadar between 1.34% and 2.97%, LinkedIn between 0.81% and 1.70%. Medium and Forbes are flatter again. For those five the engine you optimise for barely changes the answer, which is worth knowing before you split a budget three ways.

This is not a quirk of our sample. Orbit Media ran 72 prompts against four models weekly, 13,184 citations in all, and found that all four agreed on the same domain for the same question 30 times out of 1,792 question-and-domain combinations, which is 1.7%. Their unit is agreement per question and ours is the general set, but the direction is the same: the engines are not reading one web. Search Engine Land, reporting Tinuiti's Q1 2026 data, reached the same verdict, that there is no universal top source, with per-platform splits that diverge as sharply as ours do on a different query set and a different method.

The Ten Kinds of Website Each Engine Cites, Ranked

Named websites past the general set are specific to our sample. Kinds of website are not, and this is the ranking that should travel to your category. The taxonomy is the one the Sources page uses, and the examples are drawn only from the general set above.

Read the three tables together and the engines separate on how far they stray from corporate pages. ChatGPT strays furthest: corporate is 62.4% there against 84.6% on Gemini, and what fills the gap is communities at 18.8%, nearly double Perplexity's 9.9% and more than three times Gemini's 5.7%, plus the record types, reference, academic and government, at 5.2% against 3.2% and 1.3%. Gemini is the opposite, the most corporate-concentrated of the three with every other kind compressed below six per cent. Perplexity sits between them and spreads widest across the small kinds: reference at 2.0% and social networks at 1.9% are both the highest of the three, which is the Wikipedia and LinkedIn rows wearing a different hat.

The news and editorial line is the same story in every table. It is where the ranked round-ups live, the "best X" page type our published study puts at the top of content-format citations, and it holds a similar share on all three engines. What changes by engine is everything around it.

The Biggest Bucket Is Your Category's Own Websites

Corporate websites are the largest kind on every engine: 62.4% of a company's citations on ChatGPT, 75.6% on Perplexity, 84.6% on Gemini. Almost none of those sites are cited for more than one company, because a competitor's product page is only relevant to one category. Websites that appear for only one of the companies in the sample carry 63.3% of ChatGPT's citations and 72.6% of Gemini's.

In our sample that bucket holds a brokerage's competitors and the finance press that reviews them, a clinic's medical references, a startup's directory listings, a retailer's platform vendors. None of that belongs in a general list, and all of it belongs in the list of the company it was cited for. Yours will hold different sites, and they will be the larger half of your citations on every engine.

One caveat on the exact figure. How much of an engine's citation volume reads as single-company depends on how many companies sit in the sample, and the share falls as the tracked set grows. Read it as the direction, not a constant. The direction is not in doubt: the general set is the part every brand shares, and the least specific part.

How Narrow Each Head Is

The general set is small, and how much of each engine its most-cited websites explain varies by a lot. The ten websites ChatGPT cites most carry 26.9% of its citations; Gemini's ten carry 9.3%. That is a 2.9× difference in how much any top-10 list, ours or a public one, can tell you about the engine it describes.

The same shape shows up in how many websites it takes to reach half of an engine's citations: 151 for ChatGPT, 527 for Gemini. Both are a small citation pool next to the open web. But a pool of a few hundred websites is not a pool of ten, and an engine drawing from the wider one is far less predictable from any short list.

The shape is not particular to our corpus. A May 2026 academic study of Google AI Overviews, over 61,212 reference URLs from 7,479 hostnames, found the top 5 hostnames account for 20.0% of citations, the top 10 for 29.7% and the top 100 for 57.1%. A different surface, the same long tail.

Why Our Lists Differ From the Public League Tables

If you have seen a public most-cited-domains tracker, some rows here will look familiar and some will not, and the differences are worth being explicit about. Ahrefs' public trackers, built on a broad set of US queries and regenerated monthly, put Reddit at 16.8% of ChatGPT's citations as of September 2026. Ours says 17.87% of a company's citations. That is close agreement from two unrelated query sets and two different ways of counting. On Gemini the same tracker has Reddit at 28.5% against our 3.43%, and on Perplexity it has YouTube at 20.8% against our 2.59%.

Four things explain gaps that size, and none of them is that one list is wrong. The categories, which is the largest: our questions come mostly from software and business-services companies, while a tracker spanning every topic people type is dominated by consumer questions, where retailers and video carry far more weight. The counting: a public tracker ranks raw volume, and we average company by company, which is why sites that one heavy account would push up a volume list are absent here. Own-domain stripping: a public tracker keeps every brand's own pages in the count. And the surface: a consumer app and a grounded API model of the same engine need not cite identically. A most-cited list is a property of the questions asked and the way they are counted, which is the reason to measure your own rather than borrow one.

What to Do With a General Set and a Type Ranking

The instruction is not to chase 7 websites. It is to cover the shared list once, work the kinds of website each engine favours, and put most of the effort into a list only you can build.

• 01Cover the shared list once. All 7 pay on all three engines, so a presence in the Reddit threads where your category is discussed, a place in the editorial round-ups, a video, a LinkedIn presence and a Wikipedia entry are the work that pays three times over rather than once.
• 02For ChatGPT, be in the record: the entity work that makes you Wikipedia-eligible, the editorial listicles in your category, and any research you can publish under your own name. The ChatGPT playbook covers the mechanics.
• 03For Gemini, be on the platforms: a video that answers the category question, and third-party writing on the publishing platforms and business press it treats as opinion worth citing. The Gemini and AI Overviews playbook separates the three Google surfaces.
• 04For Perplexity, exist properly everywhere a company can be looked up: LinkedIn, for the people as much as the page; video; the review platforms and directories that prove you are real and rated. The Perplexity playbook explains why an engine that cites more sources per answer rewards breadth.
• 05Then map your own tail, which is most of the work. Record every URL each engine cites for your prompts, split by engine, and treat the competitor pages, trade press and communities that show up as the real list. The off-site GEO playbook sequences that work by citation share.

This is what Ranqo's Sources page shows for a brand's own prompts: the Domains tab ranks every website each engine cited for your questions, the Source Types view gives the same breakdown as the type tables above for your category rather than ours, and Outreach turns the pitchable pages into a work queue. The general set in this post is the part every brand shares. The product is for the rest.