Two in three sources AI cites are company websites

When somebody asks an AI about a product category, the model leans on a specific handful of sources to write its answer. Knowing which ones changes where the effort is worth spending.

We went through everything ten engines cited over almost three months. Here are the numbers.

What we measured

Every answer Bee LLM collected between 2 June and 22 August 2026, across all accounts:

Two things get mixed up constantly, so it is worth pulling them apart:

• Cited source. The AI links it inside the answer. The user sees it and can click it.
• Opened source. The AI read it to make up its mind but never showed it. No clicks come from it, although it does shape what the model ends up saying.

The AI reads ten pages and shows you fewer than three

On average each answer opens 10.3 pages and links only 2.7. It reads almost four times as much as it shows.

There is a second figure that gets far less attention. Only 51.5% of answers cite anything at all. The other half answers without a single link. In those cases a brand can be recommended and still not earn one visit.

Being cited and being recommended are not the same thing. You can have one without the other, and they are worth measuring separately.

Two in three citations are company websites

This is how the 56,966 citations split by type of site. We stripped out the websites of the brands we track and their competitors, so self-citation does not inflate the figure:

Forums, the source everyone talks about, come fifth and stay under 3%. Reddit on its own picks up 1,060 citations, which sounds like plenty until you set it against a total of nearly 57,000.

What dominates is far more boring and far easier to control: company pages explaining what they do. Product pages, comparisons, use cases, documentation.

Every engine decides differently

Averaging ten engines into one number hides some wide gaps. These are the three figures that separate them best:

Three things fall out of that table:

• Perplexity always cites. It is the only engine that links something in 100% of its answers. If clicks are what you are after, that is where being present pays most.
• DeepSeek and Qwen do not search. They answer from what they learned in training. There is no competing with fresh content there. Either the model already knows who you are, or you are absent.
• AI Overviews barely cites. Just over one source per answer, and only in four out of ten. It is the format that summarises most and shares least.

The ChatGPT API stands out on its own terms. It opens 36 pages on average and links three. It reads an enormous amount and shows very little.

The sources that turn up in the most projects

A domain with hundreds of citations inside a single project is a local quirk. What tells you something about the wider ecosystem is the domains that appear across different projects. These showed up in the most:

Reddit and YouTube really are everywhere. Both turn up in 8 of the 14 projects. What they do not have is weight inside each answer. Wide reach, small share.

Mainstream press performs better than its reputation suggests. El País appears in nine different projects, more than any other source in the set.

What to do with this

Four practical conclusions, in order of effort:

1. Your own site is still the main piece. Two in three citations go to company pages. Explaining clearly what you do, who it is for and how it differs still carries most of the weight.
2. Forums are not the shortcut they look like. They supply under 3% of the citations. Useful, but not where the match is won.
3. Press and directories punch above their size. Media, directories and review sites add up to 17.5% between them, and they are places you can work on deliberately rather than hope for.
4. Pick the engine to match the goal. If you want visits, Perplexity and ChatGPT web cite generously. If you want to be named, DeepSeek and Qwen depend entirely on what the model already knows about you.

One caveat about these numbers. They come from 14 projects across 7 sectors, mostly software, hospitality and professional services in Spain. Treat them as a bearing, not a universal law. Measure your own case, because the split shifts a good deal from one sector to the next.

Read next: the sources the AI reads but never cites · why track the URLs AI cites · AI cites deep pages, visitors land on the homepage.