If the Answer Is Not Built From Vendor Sites, We Told You to Go Where It Is Built. Here Is What That Turns Out to Be.

Our matched-category study found that in most categories the answer is not assembled from vendor sites, and it closed with advice: go and be placed in the sources that do hold the slots. This is what that advice is worth when you count the sources. Across 29 cells of five buying questions in six markets, 119 durable citation slots belong to somebody other than a vendor, and they are spread across 73 different domains. Fifty-eight of those, 79 per cent, appear in exactly one cell. YouTube appears in 24 of the 29.

So the layer is one video platform, a forum, one national portal per market, and a tail of fifty-eight sites that each show up once. That is not a media plan.

Disclosure: EchoWi sells AI visibility measurement, and this piece walks back the practical half of advice we published ourselves four days ago, which is a thing a seller has an obvious interest in not doing. The five questions, the six markets, the surface, the run count and the dates are in the piece it corrects, and every one of the 73 domains is named in our measurement register so anyone can recount them.

The short version

1. 119 of the 153 durable slots in the study go to non-vendors. That is the finding the previous piece leaned on, and it holds.
2. Those 119 slots are held by 73 distinct domains, and 58 of them appear in exactly one of the 29 cells. The layer is mostly one-offs.
3. YouTube is in 24 of 29 cells. Nothing else comes close: Reddit is in seven, a German founders’ portal in five, Forbes in three. That count is of the domain, and reading the same slot page by page shows it can be five different videos by five different creators.
4. Only nine domains cross a market at all. Six of the nine do it because two of our markets share a language, which is a smaller finding than it looks.
5. So “go where the answer is built” is one honest instruction and a long list of coincidences. The honest instruction is video.

What we said, and why it needed checking

The matched-category piece ends with a section on what to do, and for the categories that give vendors almost nothing it says the work that pays is placement in the sources that do hold the slots, naming magazines, forums, agencies and video.

That sentence is true and it is not actionable, which we did not notice because we had counted the vendor column and never counted the other one. A buyer reading it would reasonably ask which magazines. This is that count.

Every figure here comes from the same 30 rows: five categories, six markets, one buying question each, three runs, one answer surface. One cell returned nothing durable at all and is excluded, which leaves 29.

The count

Seventy-nine per cent of the domains that hold this layer up appear once. A vendor who wants to be where the answer is built is looking at a list where four in five entries are worth one cell in one market.

The fifteen that repeat, and how fast they fall off

The shape is the finding. One domain holds a slot in 83 per cent of the cells, the second holds a quarter of them, and by the fourth entry we are at three. Everything else is two or one.

That distribution is what makes the practical advice hard. If the layer were forty domains each holding six or seven cells, a buyer could work a list. It is one domain holding almost everything, and then noise.

The nine that cross a market, and why six of them do not count for much

Only nine of 73 appear in more than one market. Six of those nine are explained by the study containing two Spanish-language markets and two English-language ones: bbva.com, softwaredoit.es and factorial.es cross between Spain and Mexico, and forbes.com, pcmag.com and ecomm.design between the United States and the United Kingdom. 01net.com crossing from France into a British answer is the one that is genuinely odd, and one odd row is a row, not a finding.

Which leaves YouTube and Reddit as the only two domains in this study that hold citation slots across languages rather than within one.

What we would say now instead

Video is the one instruction that survives the count. A domain in 24 of 29 cells across six markets and five categories is not a placement opportunity in the usual sense, it is where a large share of this layer physically is. Our earlier work on whose videos get cited found that none of the ten cited videos belonged to a vendor and eight distinct channels covered them, so the reachable version of this is small channels rather than a platform deal.

The national portal is the second, and it is a different job in every market. Germany’s founders’ portal holds five cells; nothing plays that role in the United States, where the same slots are spread over Forbes, PCMag, the Chamber of Commerce and a long tail. That is one relationship per market, found by measuring that market.

And the tail is not work, it is weather. Fifty-eight domains holding one cell each is what a retrieval layer looks like when it is assembling an answer from whatever is available, and a plan built on it would be a plan to be lucky fifty-eight times.

Which puts the weight back on the categories where vendors do hold slots. Accounting gives vendors 14 of 37 durable slots across six markets. In a category like that a vendor site is a candidate and the ordinary work applies. In hosting, at one of 27, it does not, and the honest answer to “what do I do” is that this is not a channel you can buy your way into yet.

What this does not show

One surface, six markets, five categories, one reading each. Every limit of the study this counts applies here unchanged, and it is the same 30 rows.

A count of cells is not a count of importance. A domain holding one cell may be the source the answer leans on hardest in that cell. This counts breadth, which is the thing a media plan needs, and says nothing about depth.

The classification is ours. A domain is a vendor when its own home page names the product the question asks about, applied uniformly, and that rule under-counts suite-branded vendors. Every domain moved out of the vendor column by that rule lands in this one, so if you disagree with the rule these numbers move with it.

Nothing here is causal or a forecast. This is what one surface returned on two dates. Whether the tail is stable is a different question, and our cross-day work suggests the answer is mostly no.

And a national portal is a category we assigned by reading. fuer-gruender.de is a founders’ portal, smallbusiness.co.uk a guide publisher, soyconta.com a blog for accountants. Those readings are in the register domain by domain.

Common Questions About Who Holds Citation Slots

If the answer does not cite vendors, who does it cite?

In this study, 119 of 153 durable citation slots went to non-vendors, spread across 73 distinct domains. The distribution is extremely uneven: YouTube held a slot in 24 of the 29 cells, Reddit in seven, and 58 of the 73 domains appeared in exactly one cell.

Can I get placed in the sources an AI answer uses?

For most of them there is nothing to work with. Four in five of the domains holding this layer appear once in one market in one category, which is a coincidence rather than a placement. The exceptions are video, which appears nearly everywhere, and one national portal per market, which is one relationship per country and has to be found by measuring that country.

Does the same publisher get cited in different countries?

Almost never. Nine of 73 domains appeared in more than one market, and six of those nine are explained by the study containing two Spanish-language markets and two English-language ones. Only YouTube and Reddit held slots across languages.

Should I still try to rank my own site?

It depends entirely on the category, and that is measurable in an afternoon. Accounting software gave vendors 14 of 37 durable slots across six markets, so a vendor site is a legitimate candidate there. Web hosting gave them one of 27, so in that category the answer is not made of vendor sites and no amount of work on yours changes it.

How many runs is this built from?

Three per question, and only questions that answered all three times are in the comparison. A domain missing from one run drops out of the durable set entirely, which is a strict bar and is stated per row in the register.

Is the tail stable enough to plan around?

Our own cross-day measurement suggests not. Asking the same frozen questions two days apart, the median overlap between the sets of domains holding every run was 0.63, and both the highest and lowest rows moved when simply read again the same afternoon. A domain that holds one cell today is the least likely part of this to be there next week.