Which pages does AI cite as sources?

Mostly deep pages. Across the nearly 100,000 cited URLs we have on record, only 19.4% are homepages: the rest are guides, comparisons, articles and detail pages one, two or more folders deep. Similarweb's 2026 report points the same way, with 65% of cited URLs sitting two or three folders deep.

Then why does the traffic enter through the homepage?

Because most people never click the citation. They see the brand in the answer and go look it up. Similarweb measures that 58.8% of AI referral traffic lands on the home, and in one of our projects with Google Analytics connected the figure rises to 73%.

Should I push the AI link toward my product page?

That is not what the data says. Product and pricing pages earn the fewest citations of any format. The citation goes to content that answers the question, and the conversion happens later, when the visitor enters through your homepage. Each page has its own job.

How do I measure this on my own site?

You need both views: which of your URLs and third-party URLs the AI uses when answering, with their depth, and which pages the AI-referred traffic actually enters through. In Bee LLM the first lives in the URLs view and the second in AI Traffic, by connecting Google Analytics read-only.

AI cites deep pages but visitors land on your homepage

We looked at nearly 100,000 URLs that AI assistants cite as sources. Only 19.4% are homepages. Yet when that user decides to visit you, between 58% and 73% of the time they come in through the home. It looks like a contradiction and it is not: those are two different jobs.

Aug 11, 2026·6 min read·Our dataAlex Amigo

There is a question we get all the time: "which page of my site should I optimize so AI cites me, the homepage or the content?". The short answer is both, but for different things. The long answer is this article, and it comes from crossing two numbers almost nobody looks at together.

Which pages AI actually cites

At Bee LLM we store every URL an AI assistant uses as a source when answering our clients' prompts. At the time of this analysis that was 97,818 distinct cited URLs. We grouped them by depth, meaning how many folders the path has.

Depth of the URLs AI cites

Across 97,818 distinct cited URLs recorded by Bee LLM

Domain homepage19.4%

One folder22.1%

Two folders23.2%

Three folders12.3%

Four folders or more22.9%

Source: Bee LLM database, URLs cited as sources by ChatGPT, Gemini, Perplexity, Claude and Google's AI summaries, deduplicated. August 2026.

Four out of five citations go to an inner page. And it is not a quirk of our data: Similarweb's 2026 report on AI search, analyzed by Aleyda Solis, puts 65% of cited URLs at two or three folders deep.

It makes sense. When a model looks for something to answer with, it wants the page that answers that exact question: the guide, the comparison, the study, the detailed spec. The homepage talks about everything and answers nothing in particular. Another industry finding points the same way: reports, comparisons and reviews are the formats with the highest citation rate, while product and pricing pages get the lowest.

Now look at where people actually enter

Here comes the twist. Similarweb also measured where the traffic AI assistants send to websites lands: 58.8% enters through the homepage. In one of our projects with Google Analytics connected, the figure is even higher: 73% of sessions arriving from an AI enter through the home.

Two different jobs

What AI links versus where the visitor enters

AI citations that go to inner pages80.6%

AI referral traffic landing on the homepage (Similarweb)58.8%

AI referral traffic landing on the homepage (project with GA4)73%

Sources: Bee LLM database (citations); Similarweb, 2026 Generative AI Landscape, with Aleyda Solis's analysis (traffic); Google Analytics 4 from a project measured with Bee LLM (traffic).

How can AI link your guide while the user ends up on your home? Because most of them never click the link. They see your brand in the answer, remember the name and look you up. Or they type the domain directly. The citation legitimizes the answer; the visit arrives by another road.

One more number seals it: only 6.8% of ChatGPT answers carry web citations at all. Most of the time an AI talks about you there is no link to click. What travels is your brand, not your URL.

What this means for your site

That the "homepage or content?" question is framed wrong. Each page has a job and they are not interchangeable.

Deep pages win you the citation

Guides, comparisons, data-backed studies, documentation. These are the pages that answer concrete questions, and answering the question is what weighs most for getting cited: in an analysis of two million citations, question-content alignment was the strongest signal. Without these pages, you never enter the answer.

The homepage converts the visit

It is where the person who met you in an AI answer lands. If your home does not say in five seconds what you do, for whom and why they should believe you, the work your guide did gets wasted. The homepage does not need to win citations: it needs to close what the citations start.

In practice, this changes where you invest:

Build pages that answer whole questions. A good comparison two folders deep does more for the citation than one more design pass on the home.
Do not force the link toward your product page. It is the format that earns the fewest citations. Let the citation go to the content and sell the product at home.
Treat your homepage as a brand landing page. Whoever enters there has already seen you recommended; do not make them hunt for the value proposition.
Measure both sides separately. Which of your URLs AI cites and at what depth, and which pages the AI-referred traffic enters through. Look at only one and you will draw the wrong conclusion about the other.

The limits, which also deserve saying

Before drawing conclusions

Our data comes from our clients' prompts. Nearly 100,000 URLs is a large sample, but skewed toward sectors that measure AI visibility.
The 73% homepage entry figure is one project. We publish it because it squares with Similarweb's 58.8% aggregate, not as a universal number.
Correlation, not an experiment. Nobody has yet moved the same piece of content across depths and measured the change in citations.
Referral traffic is dominated by ChatGPT, which drives over 90% of sessions from assistants. On engines that barely refer traffic, what counts is the mention, not the click.

Questions we get

So the homepage does not matter for GEO?

It matters a great deal, but in the second half of the journey. It is not the page you win the citation with; it is the page where you collect the result of having won it.

What is the ideal depth for a citable page?

There is no magic number, but the one-to-three folder band concentrates 57.6% of the citations in our base. More than depth itself, what decides is that the page answers a complete question on its own.

How do I know if my case looks like the 58.8% or the 73%?

By connecting your analytics. The exact split depends on your brand and sector: the stronger the brand, the more people look it up instead of clicking the citation, and the more the homepage weighs.

Does this hold across all engines?

The deep-citation pattern repeats across every engine we measure, with nuances. And on the traffic side, ChatGPT dominates to the point that your referral analytics is, in practice, ChatGPT analytics.

See both sides on your own site

Bee LLM shows you which URLs AI uses when answering about your sector, with their depth and page type, and if you connect Google Analytics read-only, which pages the AI traffic enters through. Free plan, no credit card.

Start free

Sources: Bee LLM database (97,818 cited URLs, August 2026); Similarweb, 2026 Generative AI Landscape: The Evolution of AI Search, with Aleyda Solis's depth analysis; What Drives Citations, an analysis of about two million citations across 10,042 URLs. To keep reading: why Google's top 10 is not enough and half of small brands never appear in AI answers.