What kinds of pages do AI assistants cite for GEO questions? 569 links, four assistants

We measured which kinds of pages four AI assistants cited when answering 23 buyer prompts, and sorted every cited link into one of nine page types. The raw material is 120 answers collected on 14 and 15 September 2026 from ChatGPT (52 answers, signed out, web search on), Perplexity (24 answers, logged in with personalisation switched off, because the signed-out version showed a sign-in wall), Google AI Mode (34 answers, private window) and the Gemini app (10 answers). Every prompt was typed verbatim, in a fresh session, and repeated two or three times, following the protocol in our measurement guide. 33 answers listed no sources at all, 22 of them in ChatGPT; the remaining 87 answers contained 569 source links.

The prompts fall into two groups. The first is how-to questions about GEO itself, eight prompts in all: what happens if GEO does not work after 90 days, whether agencies offer guarantees, what a contract should say about results and data, whether an SEO team can handle GEO, and what a company can do before paying an agency. The second is questions where the asker is choosing a vendor: the best business bank account for a UK startup, a safe neobank for freelancers in Spain, the cheapest transfer from the UK to India, personal-injury lawyers in Miami in English and Spanish, the best GEO agency for Hindi, and the Gemini-run prompts about GEO agencies, their prices and llms.txt.

Classification followed one rule: the page type matters more than the site type. An agency that publishes a ranking of agencies with itself inside gets the directory code; a software vendor's research post gets the tool code; a bank's account page and a law firm's home page get the same code as an agency's services page, because in each case the cited company is the one selling what the prompt asks about. One classifier did the work on 22 September 2026; 116 links with unambiguous domains (Reddit, Google and OpenAI documentation, arXiv, our own pages) were coded automatically.

Across all four assistants, 271 of the 569 links (48%) led to a vendor's or agency's own site: 120 to a commercial page and 151 to an article on such a site. Directories, rankings and comparison sites came next with 90 links (16%), then tool vendors and forums with 56 each (10%), platform documentation with 39 (7%), research and government sources with 28 (5%) and media with 20 (4%). Nine links (2%) pointed at pages about a different GEO altogether.

The figure worth pausing on is media at 4%. Newspapers, trade press and industry publications were the source of one link in twenty-five. Everything else came from the sites of companies that sell something, from platforms, from lists or from people talking to each other on forums. That is one dataset, on one topic, over two days, so it is not a law; but it does not support the assumption that an assistant's answer is assembled mainly from press coverage.

The table shows the count and share per assistant. Shares are rounded to whole percentages, so columns can add to slightly more than 100%. The Gemini app column has 21 links from 4 sourced answers and is shown for completeness only.

The mix changes with the intent of the question: how-to questions are answered from articles, and vendor-choice questions are answered from vendors' own pages and from lists of vendors. On the eight how-to prompts about GEO (263 links), 42% of links went to articles on agency and vendor sites, 14% to tool vendors, 14% to forums and only 7% to a vendor's commercial page. On the vendor-choice prompts (306 links), the commercial page took 33%, directories and comparison sites 25%, and articles fell to 13%.

That split matches how the assistants describe their own retrieval. OpenAI's help page on searching the web with ChatGPT, checked on 16 September 2026, says the assistant typically rewrites a query into one or more targeted searches before answering. A how-to question becomes a search for explanations, which finds guides; a "which bank" question becomes a search for products and comparisons, which finds account pages and lists. The page type an assistant cites is largely the page type its rewritten search surfaces.

Platform documentation is the one row where a single assistant drives the whole number. Across the how-to prompts it is 6% of all links, but 25% of ChatGPT's; without ChatGPT it barely appears. For a company writing a how-to guide, the competitor for a citation in ChatGPT is developers.google.com, not another agency.

The three assistants draw on three different pools, and the differences are large enough to show at this sample size. ChatGPT is the documentation reader: 25% of its 143 links went to platform documentation, chiefly developers.google.com and help.openai.com, and it cited no forum at all. It also had the highest share of commercial pages (45%), most of them from the banking and Hindi-agency prompts, and left 22 of its 52 answers without any source.

Perplexity is the article collector. It cited a page from a vendor's or agency's own site in 23 of its 24 answers (96%), returned around thirteen sources per answer, and sent 32% of its 311 links to vendor articles, 19% to directories and 12% to tool vendors. It cited no platform documentation on any prompt. Its lists were also the most repeatable: on the in-house prompts the first and second run returned almost the same domains.

Google AI Mode is the forum reader. 22% of its 94 links went to Reddit threads, LinkedIn posts, Quora and YouTube, against 9% in Perplexity and 0 in ChatGPT; a further 34% went to vendor articles. It returned three or four sources per answer, and cited a vendor's own site in 18 of its 29 sourced answers (62%). The Gemini app, with 21 links from 4 sourced answers, leaned on Medium and tool vendors, but the sample is too small to characterise.

Perplexity also produced the one classification surprise. Nine of its 311 links, all on the what-if prompts and seven of them on the contract question, pointed at a different GEO: the prison operator GEO Group, graduate-employee unions called GEO, and legal templates found through the word "contract". That is a term-entity problem, not a retrieval bug. The acronym is young, the assistant searched the literal string, and the largest entities carrying it are not in marketing. A company whose name or category shares a string with something older should expect the same, and the fix is the entity work described below.

The most-cited domains were a documentation site, a forum and four banks, in that order, once agency domains are set aside. 221 distinct domains appeared across the 569 links, and most appeared once. The table lists every non-agency domain with seven or more links; agency and consultancy domains are omitted by editorial policy, since we do not name competitors on this site, and get-geo.ai is included because it was counted the same way as every other domain.

Two things stand out. developers.google.com was cited only by ChatGPT, on the in-house, what-if and fintech prompts alike, and reddit.com only by Perplexity and AI Mode. The four banks and fintechs reached the top of the table through the three UK and Spain banking prompts, where Perplexity in particular returned nine or ten pages from a single provider in one answer: account pages, fee schedules, comparison pages the bank published itself. seg-social.es, the Spanish social-security site, was cited twelve times in Perplexity's answers about neobanks for freelancers, mostly through lists of registered collaborators, which is a reminder that a source count is not a quality score.

Our own domain sits in the table with 12 links, 9 of them from ChatGPT. Eight of those came from the four Hindi-agency prompts, which we wrote around our own case, so the number says more about the prompt set than about the market; we show it because leaving out the one domain we control would be a worse choice.

It means the page most likely to be cited for a buyer's question is a page on your own site that answers that question, and the second most likely is a list that includes you. Half of all links in this count went to vendors' own sites, and that share was highest exactly where it matters commercially: on vendor-choice prompts, one link in three went to a company's own commercial page. An assistant looking for "the best business bank account for a UK startup" cited the banks' own account pages, fee tables and comparison pages, not a journalist's round-up.

The first consequence is one page per buyer question, in two forms. How-to prompts were answered from articles (42% of links), so a question your buyers ask in how-to form needs an article that answers it in the first paragraph. Vendor-choice prompts were answered from commercial pages (33%), so the same topic needs a services or product page that states what you sell, for whom, at what price, with the name and description used everywhere else. Our guide on how to optimise your company for AI search describes the page format; our guide on why ChatGPT doesn't recommend you covers the entity side, which is also the cure for the wrong-GEO problem above.

The second consequence concerns directories. On vendor-choice prompts, directories, rankings and comparison sites took 25% of links, and in ChatGPT, which cited no forums, lists were the main way a name entered an answer without its own page being cited. Being present in the lists an assistant already cites is therefore worth checking before you pay for anything; our guide on whether to buy listicle placements explains how to check a list first and what a placement can and cannot do. For regulated categories, the banking prompts show the same pattern with a twist, which our guide on GEO for financial services takes further: the provider's own page, a comparison site and a regulator, side by side.

The third consequence is that media coverage is corroboration, not the main event. At 4% of links, press is the rarest source type in this count. It still matters for brand mass, the layer that moves over quarters, but a plan that spends the first quarter on PR while the site has no page per question is building the slowest layer first. Our guide on whether GEO works sets the layers in order, and our guide on GEO in-house versus an agency lists what a team can do itself in the first week: the pages, the entity description and crawler access.

Finally, the three-pool difference argues for measuring per assistant rather than in aggregate. A page that ChatGPT cites because it reads like documentation may never appear in AI Mode, which is reading Reddit; a Perplexity citation says little about ChatGPT. The free protocol in our guide on how to check your AI visibility runs the same prompts in each assistant and records the sources, which is the only way to see which pool you are missing from.

The count is small, short and single-handed, so read it as a first measurement rather than a benchmark. It covers 120 answers over two days, 23 prompts written by us, four assistants and one language for most prompts. Perplexity was used logged in, with personalisation off, because the signed-out version was blocked, so its column is not a clean session by our own definition. The Gemini app column has 21 links and is not used for any conclusion. One classifier did the sorting; 27 links were classified from the URL and path alone because the site could not be opened from the classifier's environment, and the original and corrected files are both kept.

Two design choices also shape the numbers. Grouping the prompts into how-to and vendor-choice was our decision, and the fintech, legal and Hindi prompts weigh more heavily in the second group than the GEO-agency prompts do. And the Hindi prompts, written around our own case, inflate the count for our own domain, which is why we report it but do not lean on it.

The count also has an edge where local services begin. In a separate study of 72 answers to lawyer searches, run on 22 September 2026, ChatGPT cited a law firm's own site in only 5 of its 24 answers (8 of 71 links) and took the firm names from a map of local businesses shown above the text, while Perplexity cited firm sites in all 24. For a business that clients look for by city, the listing on that map can matter as much as any page type counted here; our study of which lawyers AI assistants recommend describes that layer.

The next remeasurement is scheduled for 26 October 2026, with the same prompts and the same nine types, so the mix can be compared over time. If you want to know which of these pools your own company is missing from, the free AI visibility audit is the starting point: ten prompts in one language where you are invisible today, the current answers with their sources, the specific changes that would put you in them, and a check of crawler access. Email hello@get-geo.ai with your site and the language you sell in.