No page type wins everywhere, and the published counts disagree far too much to be read as a league table. One study of what ChatGPT cites put homepages and landing pages at 23.8% of cited URLs and blog articles at 1.9%. A sector study of travel queries put landing pages at 71.2%, blogs at 24.2% and homepages at 4.6% for the same engine. Both measured carefully; they measured different populations. What moves between them is the category and the question being asked, because a how-to question reaches for reference and brand-owned material while a market question reaches for journalism. The usable version of the question is not which page type gets cited in general but which of your pages is cited for the prompts your buyers actually type, and that is a list you can pull in an afternoon.
Your homepage, your blog and your product pages all get cited, in proportions that change with the category and the question, and no published count settles it in general. The two largest public counts differ by more than twenty times on the share of citations going to blog articles, which is a good reason to distrust any confident single answer including this one.
Ahrefs exported the top 1,000 pages ChatGPT cited in September 2025 and classified them, finding Wikipedia at 29.7%, homepages and landing pages at 23.8%, educational pages at 19.4%, app stores at 6.6%, reviews at 5.8%, news and media at 5.2% and blog or article pages at 1.9%, in "ChatGPT's Most Cited Pages" by Louise Linehan (28 October 2025, read 19 November 2026). The author notes the categorisation was done by a model and should be read as directional.
SALT.agency classified 1,483 URLs ChatGPT cited and 2,279 cited in Google AI Overviews across six travel categories and found ChatGPT citing landing pages 71.2% of the time, blog content 24.2% and homepages 4.6%, with Google AI Overviews at 73.4%, 22.6% and 3.9%, in its travel sector AI visibility study by Dan Taylor (3 January 2026, read 19 November 2026).
Read together, the two say something more useful than either alone. Whichever page type dominates in a sample is a property of the questions in that sample, so what to plan around is the set of questions your buyers ask rather than the template to build.
Two page-type studies disagree because they sample different populations and classify with different rules, and neither flaw is hidden. One counted the most-cited pages on the open web across every subject, where encyclopedias and app stores crowd out everybody's blog. The other counted every citation inside one commercial sector, where the pages that answer the question are product and category pages by construction.
| Page type | Top 1,000 pages ChatGPT cited, all subjects, Sept 2025 | ChatGPT, travel queries, Jan 2026 | Google AI Overviews, travel queries, Jan 2026 |
|---|---|---|---|
| Homepages and landing pages together | 23.8% | 75.8% | 77.3% |
| Homepages alone | Not reported separately | 4.6% | 3.9% |
| Blog and article pages | 1.9% | 24.2% | 22.6% |
| Wikipedia and reference | 29.7% Wikipedia, 2.2% dictionary | Not in the sample | Not in the sample |
| Educational pages | 19.4% | Not a category used | Not a category used |
| What was counted | The 1,000 most cited URLs, any subject | 1,483 cited URLs, six travel categories | 2,279 cited URLs, six travel categories |
Category effects inside a single study make the same point. In the SALT travel data, homepages took 15.4% of Google AI Overview citations in the luxury category and 1.4% in package tours, which is a tenfold swing between two neighbouring segments of one industry measured the same way on the same dates.
Classification is the other half of the gap. A page an engine cites can reasonably be called a landing page, an educational page or a blog article depending on who is labelling, and the Ahrefs study says plainly that a model did the labelling. Two honest analysts can produce different tables from identical URLs, which is why a page-type percentage copied from a blog post into a strategy deck is usually the least reliable number in it.
None of that makes the studies useless. The pattern that survives both is that pages answering a specific question get cited and brochure pages mostly do not, whatever the label on the template. Our guide to how AI engines choose which sources to cite covers the retrieval mechanism underneath that pattern.
The question being asked decides more about which page gets cited than the page type does, and that effect has now been measured across engines. Muck Rack analysed more than 25,000,000 links cited by ChatGPT, Claude and Gemini and found that industry-trend questions cited journalism in 46% of responses, more than twice the rate for how-to and comparative questions, in "What is AI Reading?" (May 2026, read 19 November 2026).
Its description of the other direction is the part worth pinning up. When readers ask what is happening in an industry, the engines reach for news coverage; when they ask how to do something, the engines lean on reference content and brand-owned material instead. So your own pages compete best on the how-to and comparison questions, and compete badly on the market-trend questions where journalism wins.
Engine behaviour varies on top of that, which is why a single blended number hides the useful detail. The same report found that ChatGPT includes citations in 96% of responses with about five sources each, Gemini in 82% with about eight, and Claude in only 55% of responses but with about thirteen sources when it does cite. A page competing for a Claude answer is competing for one of thirteen slots in a minority of answers.
Two things follow. Build the pages that answer the questions where owned material has a chance, which covers most of the buying journey. And stop judging a product page by whether it appears in answers to questions nobody would use a product page to answer.
Page format does have a measured effect, and the largest study to test it reports the effect as modest and format-specific rather than general. Discovered Labs analysed about 2 million citation observations across 10,000 cited pages on ChatGPT, Claude, Google AI and Gemini and reports a standardised coefficient of plus 0.39 for pricing pages and minus 0.12 for listicle-style review pages, with content alignment at plus 0.37 within confidence bounds of plus 0.33 to plus 0.41 and q below 1e-73, in "What actually drives AI citations" (2026, read 19 November 2026). Discovered Labs sells AEO services, which is worth knowing when reading any of its coefficients.
Alignment carrying a larger effect than any format is the result to act on, and it is also the result that makes format advice mostly beside the point. A page whose words match the way buyers phrase the question wins against a better-formatted page that uses internal vocabulary. Our published methodology weights that finding more heavily than any other.
The same study reports smaller positive effects for page length and for similarity between the title and the prompt, and reports that schema markup and performance scores showed no significant independent effect once the domain was controlled for. Our answer on whether schema markup helps AI citations works through what that does and does not mean for a dev ticket.
So the honest instruction for a page is narrow. Answer one question the way it is asked, put the answer in the first sentence, and let the template be whatever your site already uses.
Your own pages are a minority of what engines cite in aggregate, and a majority in some categories. That gap is the last reason to distrust general page-type advice. In the Muck Rack study, owned and paid content together accounted for about 16% of citations, with earned sources at about 84%.
Our own category research found close to the opposite, in a category where buyers ask about suppliers. For our pharmaceutical process skid study we ran the buyer's question across ChatGPT, Google AI Overviews, Perplexity and Claude, five times each, for eight skid categories plus two regional markets, and found 73% to 89% of citations pointing to fabricators' own websites rather than to directories or listicles, consistent across six categories, with no single fabricator named by all four engines in any contestable category.
Both results can hold because the samples answer different questions. A general prompt set spanning consumer subjects reaches encyclopedias and journalism; a prompt set made of supplier-selection questions in a specialist industrial category has nowhere to go except vendor sites, because nobody else has written the material. Where your category sits on that line decides how much your own pages can be expected to do.
Retrieval remains the larger term either way. Chen and Koudas found that which pages the engine's own search returns explains about 32% of citation variation against about 2.8% for page properties, across 158,099 answers and 468,037 citations, in "What Drives AI Citations?" (arXiv 2610.15033, 10 October 2026). Being the right page type does not help a page that is never in the retrieved set.
Finding out which of your own pages gets cited takes a prompt set and a log, and it beats every published page-type table because it describes your category rather than somebody else's. Write the twenty questions your buyers actually ask, run them on each engine you care about, and record the prompt, the engine, the date and every cited URL.
Classify the cited URLs yourself after that, using labels that mean something for your site: homepage, product or service page, comparison page, documentation, blog article, third-party page. A pattern usually appears within two runs, and it is often not the pattern anybody expected. Pages nobody has touched for two years show up, and the new content hub does not.
Treat small numbers with suspicion before acting on them. BrightEdge's volatility analysis found frequently cited domains moving about 0.7% week to week against swings above 50% for sporadically cited ones, with an inflection near 50 citations where weekly volatility drops from 50% to 8%, in "AI Search Engine Citation Volatility: The 70x Stability Gap" (23 October 2025, read 19 November 2026). Below that volume, one missing citation is a sample rather than a trend.
If you would rather see the list than build the pipeline, a free visibility assessment produces it from outside your systems: your buying questions sampled across ChatGPT, Perplexity, Google AI Overviews and Claude, with the cited URLs behind each answer, plus a read of your public pages as a crawler sees them. Our guide on how to track AI visibility covers how to keep the series running afterwards.
Marketing teams deciding where to put their content budget ask these five questions most often about which pages AI engines cite.
Sometimes, and the published share swings widely by category. One classification of the top 1,000 pages ChatGPT cited put homepages and landing pages together at 23.8%, while a travel sector study put homepages alone at 4.6% of ChatGPT citations, and inside that same study the luxury category reached 15.4% in Google AI Overviews. A homepage gets cited when it is the only page that answers the question, which is common for small sites and rare for large ones.
Not reliably, and the two largest counts disagree about it. Blog or article pages were 1.9% of the top 1,000 pages ChatGPT cited in the Ahrefs classification, while blog content was 24.2% of ChatGPT citations in the travel sector study, where landing pages took 71.2%. What the two agree on is that pages answering a specific question get cited and undifferentiated pages do not.
Usually yes for questions with distinct answers, and no when two questions have the same answer, because two of your pages competing for one answer tends to produce neither. The test is whether the first sentence of each page would differ. If it would not, the second page is a section of the first.
Frequently for how-to and implementation questions, which is where Muck Rack found engines leaning on reference and brand-owned material rather than journalism. Documentation that is crawlable, server-rendered and outside a login is doing the same job as an article, and in technical categories it often outperforms the marketing site.
No, because the page type that gets cited most is a property of the sample rather than of your category, and the measured page-level effect that beats every format is vocabulary alignment with the question. Rewriting a page to look like a listicle is work with no measured return; rewriting its opening sentence to answer the question a buyer actually types has one.
Related guides
A free visibility assessment samples your buying questions across ChatGPT, Perplexity, Google AI Overviews and Claude, and reads your public pages the way a crawler sees them.
Request a free visibility assessment →