Published September 8, 2026
41% of what Google's AI Overview cites is YouTube or Reddit
Ahrefs traced the sources behind Google's AI Overviews across 3 million US searches. YouTube and Reddit together take 41% of the citations; the top three sources take more than half; the biggest shop in the top ten is Amazon at 3.3%. Why an AI answer reads where people talk rather than where you sell, what that means for a product page, and what to do.
When Google shows an AI Overview at the top of a search, it writes the answer from a handful of pages and cites them. Ahrefs, whose Brand Radar tool tracks those answers, looked at more than 3 million US searches across every topic and counted which domains got cited. Two of them, YouTube and Reddit, account for 41% of all citations. Add Facebook and the top three take 51.5% of everything the AI Overview cites across its 50 most-used sources. (Ahrefs, The 50 Most-Cited Websites in Google AI Overviews, September 2026)
For anyone who sells something, the interesting line is further down the table.
The top ten
| Rank | Domain | Share of citations |
|---|---|---|
| 1 | YouTube | 22.9% |
| 2 | 18.5% | |
| 3 | 10.1% | |
| 4 | 8.8% | |
| 5 | 5.6% | |
| 6 | Quora | 4.7% |
| 7 | Wikipedia | 4.0% |
| 8 | TikTok | 3.3% |
| 9 | Amazon | 3.3% |
| 10 | Walmart | 1.0% |
Nine of the ten are places where people talk, post, ask and answer. One is a reference work. Two are shops, and the biggest of them, Amazon, gets 3.3% of citations: one seventh of YouTube. Every other shop on the internet shares what is left below the top 50.
Ahrefs measured Google's AI Overview specifically, on US searches. Its earlier count for Perplexity showed the same shape, Reddit and YouTube at the top, Amazon the only large retailer near it. ChatGPT is not in this dataset and differs in its habits, but the pattern of reading community sources first shows up wherever anyone has counted.
Why an AI answer reads where people talk
An AI Overview is not written from a memory of the web. When a question comes in, the system searches, retrieves a set of pages, and writes from them. Which pages it retrieves is decided by two things: what ranks for the searches it runs, and what kind of page tends to contain an answer in the form a model can use.
Community pages win on both. A Reddit thread titled "which espresso machine under 500 is actually good" is the question itself, in the words a buyer uses, followed by twenty answers with reasons, comparisons and caveats. A YouTube review has a transcript that walks through the product for ten minutes. Both are dense with the sentences a model needs to write a recommendation: X is better than Y for Z, because. A product page, by contrast, is one side's description of one product, written to sell it, often with the comparable facts in a spec table or an image. To a model assembling a shortlist it is a source about one option, not a source for the decision.
So the AI reads where the deciding happens, and cites the pages that already contain comparisons. Your product page is read, if at all, to confirm a fact after the shortlist exists.
What it means for a product page
Three consequences, in order of how uncomfortable they are.
Your page is rarely the reason you get named. If a buyer's question is "which one should I buy", the answer is built from reviews, threads and videos. Whether your product appears depends on whether it appears there, said by someone who is not you.
When your page is read, it is read for facts. Price, stock, specifications, shipping, returns, compatibility: the things a model checks before it repeats a claim from a review. If those facts are missing, in an image, or contradict the review, the model has nothing to confirm and moves on to a product it can.
The shops that do get cited are the ones that host the conversation. Amazon is in the top ten not because it is a shop but because its pages carry thousands of reviews and Q&A, the same kind of text as Reddit. A shop with a review section that is real, long and specific is, to a model, a small community site.
What to do
None of this is a reason to give up on your own site. It is a reason to stop treating the product page as the only place the decision is made.
- Be discussed where the AI reads. Find the threads and videos that answer your category's buying question and make sure your product is in them, honestly: sample units to reviewers, answers in the threads, a real presence in the community that talks about what you sell. A single well-reasoned Reddit answer that names your product is worth more, to an AI Overview, than another paragraph of copy.
- Put the comparable facts on your page as text and structured data. Price, availability, model year, the three specifications the category is compared on. This is what the model reads when it comes to confirm, and it is the cheapest change on this list.
- Host the conversation yourself. Long, specific, unedited reviews and a Q&A section make your page the kind of page that gets cited, not just checked.
The question on the card is the one to answer first: where was your product last discussed by someone who is not you? If the answer is "nowhere recent", that is where the AI's reading list, and your absence from it, comes from.