Not One of the 91 Pages an AI Overview Cited Was a Homepage

On 14 August 2026 we asked ten small-business buying questions across two markets and recorded the exact URL of every source an AI Overview cited, not just the domain. The ten answered cells printed 92 citations that resolve to 91 distinct pages across 73 domain slots. Nine of the ten had at least one domain supplying more than one page. And not one of the 91 pages was a homepage.

Every citation study in this field, ours included, counts domains. This one counts pages, and the two do not say the same thing.

Disclosure: EchoWi sells AI visibility measurement, and this study narrows one of our own published findings rather than supporting it. The method is a single read per question, ten cells on AI Overview and seven on AI Mode, all on one day, with the full URL of every citation recorded. The counts are in our measurement register, every question is printed below, and anyone can repeat it against us in an afternoon.

Correction, 14 August 2026: an earlier version of this piece said the CRM cell cited five different videos by five different creators. It printed five links, which resolve to four videos from three channels, because one video is linked twice and one channel published two of them. Every count below now separates citations from pages, which is the same distinction this piece makes about domains, and the totals moved from 92 pages to 92 citations and 91 pages.

The short version

1. Ten answered cells printed 92 citations resolving to 91 pages from 73 domains. A domain-level count, which is what every study in this field reports, cannot see 18 of those pages.
2. Nine of the ten cells had a domain that supplied more than one page. The widest printed five citations, and those five links were four videos.
3. In seven of those nine the domain was YouTube. In the others it was Zapier, Gusto and Sage, so this is not a video story alone.
4. Not one of the 91 cited pages was a homepage. Every single citation pointed at a deep page. On AI Mode, measured afterwards in two markets, three of 120 citations were homepages.
5. This narrows a finding we published ourselves. We have written that YouTube holds durable slots in 24 of 29 category cells. That is true of the domain. At page level, in one cell those five links are four videos from three channels, which is a different thing to sell against.
6. It is seventeen cells across two surfaces, not a rate. One draw per cell, and we explain below why repeating the draw would not have helped.

Why everyone counts domains, including us

This is not an oversight anybody made. It follows from the instrument.

If you want to state how often something is cited, you have to repeat the question, because assistants re-plan retrieval on every request and a single answer cannot tell a stable source from a coincidence. The tools that repeat a question and bypass caching report back at the level of the domain. So every figure in this field that carries a denominator, ours included, is a figure about domains.

That has been good enough for the questions we were asking. Whether a category answers from a fixed set, whether that set survives a reworded question, whether it survives two days, whether it survives a change of surface: all of those are properties of a set of sources, and a domain is a reasonable name for a source.

It stops being good enough the moment somebody asks what to do about it. “Reddit holds a slot in your category” is a fact. It is not an instruction. The buyer’s next question is which page, and every study in this field, ours included, has been answering a question they did not ask.

What one draw can and cannot say

The honest limitation sits at the front, because it decides how the rest should be read.

Repeating this call does not buy a second observation. The cheap call that returns full URLs serves from a cache. We re-verified that during this very sweep and it is worth reporting, because it is the kind of thing a buyer should ask any vendor: we asked the CRM cell twice with two different strings, “what is the best CRM software for a small business?” and “What is the best CRM for a small business?”, and the two came back with a byte-identical citation list, the same eleven URLs in the same order. Two different questions, one cached answer.

So there was no point running each cell three times. What we did instead is what the register already does on another axis: raise the count of questions rather than the count of runs. Ten questions in two markets is ten observations of the structure, and structure is what this study is about. Nothing below is stated as a rate, and no percentage in this piece has a denominator made of repeated runs.

One cell returned nothing. The Spanish password manager question produced no AI Overview at all on the day. That is recorded in the register as a cell that was asked and did not answer, named in the table rather than rendered as a row of zeros, and it is excluded from every counter here. A surface that did not render is a different fact from a surface that ran and cited nobody, and collapsing the two is how a study invents a finding.

The numbers

Ten answered cells, all AI Overview, eight in the United States in English and two in Spain in Spanish. The middle column counts citations rather than pages: the CRM row prints eleven links that resolve to ten pages, and it is the only row where the two differ.

Two things are worth reading off that table before anyone builds an argument on it.

The first is the gap between the two middle columns. Across the ten cells it is 91 pages against 73 domains, so 18 pages are invisible to the way this is normally measured. That is a fifth of the evidence the answer was actually built from.

The second is the last column. The United States ecommerce cell is the one where every domain supplied exactly one page, and it would be easy to read that as a property of that category. It is not: the same category in Spain has one domain supplying four. One cell is one cell, in this study as in every other.

The carousel, and what our own finding actually says

We have published that YouTube holds a durable citation slot in 24 of 29 category and market cells. We stand behind the measurement. This study changes what it licenses you to do.

At the level of the domain, YouTube holds the slot. At the level of the page, the CRM cell prints five YouTube links, and they are not five videos: one video is linked twice, once with a start time and once without, so the five links are four videos. Opening each of the four, they come from three channels, because Consumer Desire has two of them alongside Merchant Maverick and Stewart Gauld. Nobody owns that slot. It is not a position a single video can take and hold, it is a rotating shelf that the platform keeps stocked with whoever made the most useful video that month.

That distinction changes the advice. “YouTube holds the slot in your category, so make a video” implies a winnable position. What the page layer shows is closer to “this category’s answer draws on a pool of independent video reviews, and to be in it you have to be one of the better ones, repeatedly.” Those are different budgets and different odds, and only the second one is supported.

It also cuts the other way, in favour of the smaller player. Five slots filled by five creators is five chances, not one incumbent to displace. The pool in this cell contains no vendor at all.

Not one homepage, in either market

This is the cleanest result in the study and the easiest to act on. Across all ten cells, 91 of 91 cited pages were deep pages. Not one was a site’s front door. Spain replicates it on its own two cells, 19 of 19.

Vendors do get cited. Gusto’s payroll product page and Sage’s Spanish accounting pages are in there, and Salesforce, Zoho and Constant Contact appear on their own domains too. Which vendor pages get cited, and for which questions, splits by what the question asks: a brand’s own site is a far more reliable source about its own price than about its category. But every one of those citations is a page written to answer something: a product page for a named product, a comparison post, a category guide. The homepage, which is where most brands put their strongest positioning statement, was cited zero times out of ninety-one.

The practical reading is narrow and firm. If you are trying to be the page an assistant reaches for, the asset is a page that answers the question in the buyer’s words. A homepage answers “who are you”, and nobody asked that.

Two ways a domain count misleads, in opposite directions

The 18 hidden pages are one direction: a domain count understates how many distinct pages the answer drew on.

There is a second and opposite one in the same data. In two of the ten cells a single publisher appeared as two different domains, because it publishes country editions on separate hostnames. The password manager cell cites both the main PCMag site and its Australian edition; the United States accounting cell cites the main site and the Middle East edition. Counted by domain, that is two independent sources agreeing. It is one publisher.

So a domain count understates page diversity and overstates publisher diversity at the same time, and the two errors do not cancel out because they land on different rows. Anybody building a concentration figure out of domain counts, and that includes every such figure we have published, should say which of the two they are exposed to.

The other surface cites one

This piece went out saying AI Overview only, and that nothing here should be carried to AI Mode without measuring. So we measured it, and then measured it again in a second market: seven cells across the United States and Spain, one AI Mode draw each, 120 citations.

AI Mode cites homepages. In the United States payroll answer, patriotsoftware.com appears twice, once at a product page and once at the root of the site, and the Spanish hosting answer cites two more roots. That is three homepages in 120 citations, against zero in 91 on AI Overview.

The seven cells, so the two that hold the homepages can be seen rather than described:

They are not spread evenly. All three sit in two of the seven cells, and both of those are the categories where the buyer is choosing a supplier rather than a platform. Two of the three are a vendor’s front page. The third is a comparison site whose front page is the ranking itself, so it is a homepage in the URL and an article in every other sense, which is worth knowing before treating the count as three brands winning a slot with their front door.

So the finding is strong on both surfaces and absolute on neither. A homepage can be cited. It is rare enough that building your case on one is still the wrong bet, and common enough that “never” is the wrong word, and the difference between those two sentences is the reason to measure the second surface rather than assume it.

The four United States questions carry one more thing worth having, because both surfaces were read on the same day with the same wording: AI Overview printed 43 citations for them and AI Mode printed 76. The second surface draws on roughly twice as much, which matches what this register has measured on other axes, and it means a page absent from one is not thereby absent from the other.

What this does not show

It is not a rate. Seventeen cells across two surfaces, one draw each. We cannot say how often a homepage gets cited, only that in ninety-one consecutive opportunities it did not happen.

The surfaces differ, and we measured the second one. The ninety-one figure is AI Overview. AI Mode, on seven cells in two markets, cited three homepages in 120 citations. ChatGPT and Gemini are still unmeasured here, and nothing should be carried to them.

Selection runs through the middle of it. These are our ten questions, in our two markets, in the small-business software category. A different question set produces a different 91.

It is descriptive, not causal. Nothing here shows that having a deep page causes citation. It shows what the cited pages were.

What would actually answer the next question

The obvious follow-up is whether a page keeps its slot, and this instrument cannot answer it, which is worth saying plainly rather than implying otherwise. Page-level durability would need repeated uncached draws that return full URLs, and the call that bypasses the cache reports domains. That is a gap in the tooling of this whole category, not a gap in this study, and until somebody closes it, every claim you read about “owning” a citation slot is a claim about a domain.

What can be done cheaply, by anyone, is the version of this on your own category: take the five questions your buyers actually ask, read one answer for each, and write down the URLs rather than the domains. If your category looks like ours, the list will contain no homepages, and it will tell you which page of yours would have to exist to be on it.

Common Questions About Page-Level Citation

Do AI Overviews cite homepages?

In this study, never. Across ten buying questions in two markets, 91 cited pages were all deep pages and none was a site homepage. That is ninety-one consecutive opportunities rather than a measured rate, and it covers AI Overview in one product category, with AI Mode measured separately above, but the direction is consistent and it matches what the format is for: the assistant is looking for a page that answers a question, and a homepage answers a different one.

How many pages does one domain get in an AI Overview?

Usually one, sometimes several. Across ten cells, 92 citations resolved to 91 pages from 73 domain slots, so on average a domain supplied slightly more than one page. Nine of the ten cells had at least one domain supplying two or more, and the widest single case was one domain printing five links that turned out to be four videos.

Why do citation studies report domains instead of pages?

Because of the instrument. Stating how often a source is cited requires repeating the question with the cache bypassed, and the tools that do that report at domain level. The cheaper call that returns full URLs serves from a cache, so repeating it returns the same answer. We verified that during this study: two differently worded CRM questions came back with an identical citation list.

Does this mean YouTube is not worth pursuing?

It means something more specific. YouTube holds a citation slot at the domain level in most of the category cells we have measured, but in the cell we examined page by page, that slot was five links, four videos and three channels. It is a pool rather than a position, so the realistic goal is being one of several good videos in a category, not making one video that owns a slot. Notably, no vendor appeared in that pool at all.

Do vendors get cited on their own sites?

Yes, but not on their front pages. In the ten AI Overview answers, vendor domains appear through product pages and category guides: a payroll product page, a Spanish accounting product page, a small-business comparison post published by a vendor. The pattern is that the cited page is written to answer a question, whoever owns it.

How would I run this for my own category?

Take the questions your buyers actually type, ask one of them per assistant answer, and record the full URL of every source rather than the domain. Count how many distinct pages each domain supplied, and check how many cited pages are homepages. It costs an afternoon and it tells you which specific page of yours would have to exist to be in the answer, which a domain-level report cannot.