Running Shoes: Zero of Seven Sources Sold Shoes. Project Management: Seven of Eleven Sold Software.

We asked which running shoes suit beginners, and Google’s AI Overview cited seven sources. Not one of them sells shoes. Four are independent shoe review sites and three are YouTube videos.

We asked which project management software suits a small team. Eleven sources, and seven of them sell project management software or something adjacent to it, including one vendor’s own comparison of itself against two competitors.

Same surface, same market, same afternoon. The difference is who was available to answer.

How this was measured: Google AI Overview, United States, English, 7 August 2026, one run per question, reading the citations the answer carried and nothing else. A single run is a draw and not a rate, which is why this article counts source types rather than reporting frequencies. Classifying a domain as “vendor” or “reviewer” is our judgement, and every domain is named below so you can disagree with it. The rows are in our measurement register, with the question and the date on each.

The three source lists in full

“What are the best running shoes for beginners?” Seven citations.

Vendors: zero. No Nike, no ASICS, no PUMA, despite all three being named in the answer text.

“What is the best project management software for a small team?” Eleven citations.

Vendors or vendor-adjacent: seven of eleven.

“What is the best email marketing platform for ecommerce?” Nine citations.

Vendors: five of nine. Independent reviewers: one.

The pattern, and why it matters more than it looks

A category either has an independent review layer or it does not.

Running shoes has four separate publications whose entire business is testing shoes. Retrieval finds them and stops. Nobody needs to ask Nike what the best running shoe is.

Project management and email marketing have almost none. What exists instead is a large body of vendor-published comparisons, several of which rank the publisher first. So retrieval assembles the answer from the people selling the products, plus whatever forum threads it can find.

This is the mechanism we guessed at in our stability study and could not demonstrate. That piece found running shoes to be the most stable category we had measured, with five domains all cited in every run, and project management among the least stable, with two of twenty. It proposed that instability tracks the number of roughly interchangeable pages competing to answer.

Here is what those pages actually are. The stable category is answered by a handful of independent reviewers. The unstable one is answered by a crowd of vendors describing themselves. Vendor comparisons are interchangeable in exactly the way the hypothesis needs: many pages, similar claims, similar authority, no reason to prefer one.

It also generalises a finding we thought was specific to our own market. We counted twenty-two of twenty-seven AI visibility tool comparisons being published by a company selling one, and thirteen US agencies each publishing a ranking that put itself first, and treated both as symptoms of a young category. They are not. Drip ranks itself best overall for ecommerce email. Sequenzy is first on its own list of twenty-one. Mailchimp’s cited page asks what makes Mailchimp the best. The self-ranking listicle is the dominant format of the cited web, not a quirk of the GEO market.

Which YouTube videos actually get cited

This was the question we set out to answer, having found in a previous count that YouTube appeared in ten of eleven measurements, since extended to eleven of twelve, without ever checking what the videos were.

Five YouTube citations across the three answers:

All five are comparison videos, and all five carry the same title formula as the written listicles around them. “Best X for Y”, frequently with the year. Not one is a brand channel showing its own product.

Two things we could confirm and one we could not.

The channels we could identify are independent. The running shoes Short belongs to RadDadBodTV; the email marketing video belongs to EmailTooltester. Neither sells the products being ranked. We could not identify the channel behind the other three, because YouTube renders its page in the browser and our fetch returns only the page shell. Three of five are therefore unverified and we are not going to assume.

Audience size looks close to irrelevant. The two videos whose view counts were visible had 7,400 and 3,600 views. The first is a 53-second Short, and Google’s AI Overview cited it alongside RunRepeat and Solereview, publications with far larger audiences. If view count is a ranking factor here, these numbers do not show it.

And one publisher got two slots. EmailTooltester was cited twice in the same answer: once for its website and once for its YouTube video. Same reviewer, same expertise, two formats, two citations.

What we would do with this

Find out which kind of category you are in before you plan anything. Read the sources behind your own buying question. If they are independent reviewers, your job is to get reviewed. If they are vendors ranking themselves, you are in a crowd and the crowd is why the answer keeps changing.

Publish the comparison video, not the product video. Every cited video here was a comparison in the same shape as the written listicles. A demo of your own product matches nothing that got cited.

Consider the second format before the second article. EmailTooltester occupied two of nine slots in one answer by publishing the same expertise twice, once as a page and once as a video. That is a cheaper route to a second citation than writing another page.

And do not read a single run as a rate. This is one run per question. It tells you what kinds of source answered, which is stable enough to reason about, and it does not tell you how often. How many runs a category needs is a separate measurement and the answer varies from about three to about thirty.

What this does not show

• One run per question. Three answers, twenty-seven citations. Enough to read the composition of each source list, not enough for any frequency claim.
• Three categories, one surface, one market, one day. Google AI Overview, United States, English.
• The vendor classification is a judgement. We named every domain so you can make your own. Reasonable people will argue about zapier.com and bigcommerce.com, which is why they are labelled adjacent rather than counted silently.
• Three of five YouTube channels are unverified. We report the two we could confirm and say so about the rest.
• View counts were visible for two videos only, so “audience size looks irrelevant” rests on two observations and is written as an observation, not a finding.
• Nothing causal. We did not publish anything and measure a change. This is a reading of what was there.

Common Questions About AI Answer Sources

Does Google’s AI Overview cite vendors about their own products?

Frequently, in categories without an independent review layer. Of eleven sources behind “best project management software for a small team”, seven sold software in or adjacent to that category, including one vendor’s comparison of itself against two rivals. Of seven sources behind “best running shoes for beginners”, none sold shoes.

What kind of YouTube video does AI cite?

Comparison videos. All five YouTube citations we examined carried the same “Best X for Y (2026)” formula as the written listicles beside them, and none was a brand channel showing its own product. The two channels we could identify were independent reviewers rather than vendors.

Do you need a large YouTube channel to be cited?

The evidence here says no, though it is thin. The two cited videos whose view counts were visible had 7,400 and 3,600 views, and one of them is a 53-second Short cited alongside review publications with far bigger audiences. That is two observations, not a rule.

Why are some categories answered by reviewers and others by vendors?

Because the reviewers exist in some categories and not others. Running shoes supports at least four publications whose whole business is testing shoes. Project management and ecommerce email have very few, so the pages available to answer are mostly written by the companies being compared.

Does this explain why some AI answers are unstable?

It is the most plausible mechanism we have found. Our stability study found running shoes to be the most stable category measured and project management among the least, and proposed that instability tracks how many interchangeable pages compete to answer. Vendor self-comparisons are interchangeable in precisely that way, and the stable category is the one answered by a few independent reviewers.

Should my company publish its own comparison listicle?

It is clearly a format that gets cited, since vendors made up most of the sources in two of these three categories. It is also the format that makes a category unstable, so you would be joining a crowd whose members each appear intermittently. Being reviewed by somebody independent is the scarcer and, on our data, the more durable position.