In Four Categories Out of Ten, Not One Durable Citation Belongs to a Vendor

We asked eleven “what is the best X” questions on AI Overview, three times each, in the United States in English, and kept only the domains cited in every single run. Eighty-four of those durable slots came back. Twenty-four of them belong to a company that sells the product being asked about. In four categories, the number is zero.

Counted across every category here rather than one at a time, fifty-eight of the sixty-seven domains holding a durable slot hold it in exactly one category, so there is no shared destination to be published on. The names missing from that list are the interesting part. No Mailchimp in email marketing. No Shopify in ecommerce platforms. No NordVPN or ExpressVPN in VPNs. No 1Password or Bitwarden holding a stable slot in password managers. These are not obscure companies losing to better-known ones. They are the category leaders, absent from the stable core of the answer to their own category’s buying question.

Disclosure: EchoWi sells AI visibility measurement, so a study concluding that you should measure your category before investing in content is a study with an interest. Here is everything needed to run it against us: the eleven questions are printed below in full, the surface is Google’s AI Overview, the market is the United States in English, each question was asked three times, and the date is 10 August 2026. Every durable domain is listed by name, here and as a row in our measurement register. Anyone can repeat this and publish a different answer.

The short version

1. Across eleven categories, 84 domains were cited in every run of their question. Twenty-four of them, fewer than three in ten, belong to a company selling the product.
2. In four categories the vendor count is zero. Project management, password managers, ecommerce platforms and web hosting produced a stable core with no vendor in it at all.
3. One category produced no stable core whatsoever. For VPNs, no domain survived all three runs. That is a different finding from “no vendor survived”, and it is counted separately.
4. The share varies enormously by category and does not follow the obvious lines. Help desk software is 83 per cent vendor-held. Web hosting is zero. Both are B2B software bought by the same person.
5. The mechanism is the question, not the vendor. Asked “what is the best ecommerce platform”, Shopify holds nothing. Asked “how much does Shopify cost per month”, on the same surface on the same day, www.shopify.com is cited. That pair was one brand when it was written here, and it has since been run over six brands in six categories: six of six were cited on the price question and one of six on the category question. The re-measurement also moved this row, from three runs of three on 10 August to two of three on 11 August, which is what a three-run “every run” is worth.

What was measured, and what “durable” means

Each question was asked three times. A domain counts as durable only if it appears in every answer that came back. A domain cited twice out of three is discarded, and so is a domain cited once.

That threshold is doing real work, and it is worth saying why rather than presenting it as neutral. An AI answer re-plans its retrieval on each request, so a single reading cannot tell a source the engine relies on from one it happened to reach for. Requiring presence in all runs is the cheapest available filter for that difference. It is also unforgiving: with three runs there is no tolerance, and one absence removes a domain entirely.

So the counts below are deliberately conservative. They undercount rather than overcount, and a vendor that appeared twice is treated as absent.

The other decision is what makes a domain a vendor. A vendor here is a company selling the product the question asks about. www.xero.com in accounting software is a vendor. customcpa.ca, an accountancy practice, is not, because it sells services rather than the software. slack.com is not a CRM vendor even though Salesforce owns Slack. salesforce.com is counted as a vendor three times, in CRM, in email marketing and in help desk, because it sells a product in each.

Those calls change the number, so every durable domain is published by name below. Recount them differently and you get a different figure from the same data, which is the point of printing them.

The eleven categories

Help desk was asked three times and answered twice, so its denominator is two runs rather than three. The rest answered all three. A run that fails is not a run that returned nothing, and the denominator is always what answered.

The four zero rows are the finding. In project management, the stable core is zapier.com, www.forbes.com, www.ifourtechnolab.com and project-management.com. In password managers it is YouTube, Zapier, the New York Times, Wired and PCMag, the last of these occupying three of the seven slots through its American, British and Australian domains. In web hosting it is Reddit, CNET, YouTube, Forbes and technologyadvice.com. In ecommerce it is Reddit, YouTube, Zapier and ecomm.design.

Not one of those fourteen domains sells project management software, a password manager, hosting or an online store platform.

The names that are not there

A list of who holds the slots is less striking than a list of who does not.

Email marketing produced the largest stable core of any category, twenty-three domains cited in all three runs. Mailchimp is not among them. Neither is Klaviyo, Brevo, ActiveCampaign or ConvertKit. The vendors that did hold slots are moosend.com, systeme.io, mailerstack.com and quickmail.com, alongside Salesforce. The rest of the twenty-three are agency blogs, affiliate comparison sites and a WordPress theme vendor.

Ecommerce returned four durable domains across two separate sessions and Shopify was in neither.

VPNs returned no durable domain at all. The nearest misses, at two runs out of three, were CNET, PCMag’s UK domain, thebestvpn.com, top10vpn.com and a vegan travel blog. No VPN provider reached even that. That is still an answer with unstable sources rather than no answer, which is a distinction worth holding onto: asked which local business is best, the same surface produces no overview at all, and a report that counts both as zero citations is describing two different worlds with one number.

Password managers put Bitwarden at one run out of three and no other provider anywhere.

The pattern that connects them is not size and not quality. It is that the question asks for a comparison, and a page belonging to one of the things being compared is a poor source for it.

The control, which is where this stops being a pattern and starts being a mechanism

Everything above is consistent with several explanations. Vendors might be blocked from crawling. Vendor pages might be thin. The engine might prefer editorial writing in general.

So we asked the same vendor a different kind of question, on the same surface, in the same market, on the same day.

Shopify is reachable, indexed and citable. When the question is about Shopify, Shopify is cited every time. When the question compares its category, Shopify is not in the answer at all.

That contrast rules out the boring explanations and leaves a specific one: a comparative question selects against the pages of the things being compared. Not because anyone penalises them, but because a page that says “we are the best” cannot be the source for “which is best”, and a page that compares can.

It is worth noting what the control does not show. Even on its own pricing question, Shopify holds one of four durable slots. The other three are commerce-ui.com, www.stylefactoryproductions.com and pagefly.io, none of which is Shopify. Winning the question about you is not the same as owning it.

The same categories in Spain and Germany, because one market is not a finding

The largest limit on everything above is that it is one market in one language. Rather than disclaim it, we ran 7 of the categories again in Spain in Spanish and the same 7 again in Germany in German, the same way. Every category in the table below is measured in all 3 markets.

Three results come out of this, and they do not all point the same way.

The ordering travels. Accounting and help desk software keep vendors in their stable core wherever they are asked. Hosting, password managers and ecommerce keep them out or nearly out. A company selling accounting software has a route through its own pages in every market measured, and a hosting company does not have one in any of them.

One category is the same in all three markets and it is the strongest result here. Web hosting produced thirteen durable slots across the United States, Spain and Germany, in three languages, and not one belongs to a hosting company. The stable core is Reddit, CNET, Forbes and technologyadvice.com in America; YouTube, Reddit and a Madrid marketing school in Spain; heise.de, experte.de, fuer-gruender.de and websiteplanet.com in Germany. Different sites every time, same kind of site, and never the vendor.

The zero does not travel by itself. Spanish password managers give durable slots to Bitwarden and Keeper while American and German answers give them to nobody, and opentiendas.com, a Spanish ecommerce platform, holds one where Shopify holds none. So a zero is a fact about a market and a category together. Two of three markets agreeing is what makes password managers a low category; the Spanish reading is the outlier, and one market alone could not have told you which one it was.

And underneath, nothing is shared. Take YouTube and Reddit out, since global platforms hold slots everywhere and their presence is not evidence of anything, and count the sources that appear in a category’s durable core in more than one market. Across all 7 the answer is 1, eesel.ai, which holds a help desk slot in both Spain and Germany. Nothing else survives a change of country, and that single name is the whole of it. This paragraph said none when the table held five categories, and widening it to seven is what produced the exception. PCMag, Wired and the New York Times do not appear in Spain or Germany; Xataka does not appear outside Spain; Chip, PC-Welt and Stiftung Warentest do not appear outside Germany.

The practical reading is uncomfortable for anyone selling a report. A category benchmark computed in the United States tells you the shape of your Spanish or German market and none of its names, and the part that looks most quotable, a clean zero, is the part least likely to survive the trip.

One note on method that changed a number here. Four of these domains were classified by reading their own homepage rather than by recognising the name, and two of those four came back different from the first guess. opentiendas.com describes itself as an all-in-one platform for building online stores, which makes Spanish ecommerce a one rather than a zero. n2f.com sells expense reports rather than accounting software, so it is not counted as an accounting vendor. Both would have moved a published figure, and neither would have been caught by looking at the domain and thinking it looked about right.

What this means if you sell something

The useful consequence is not that content marketing fails. It is that whether your own pages can hold a citation slot is a property of your category, and you can find out in an afternoon instead of a year.

Ask your category’s buying question three times. Write down the domains that appear every time. If none of them sells anything, your own site is not the asset that wins that question, and a year of writing “why we are the best CRM” will not change it. What wins it is being the subject of the pages that do hold the slots, which is a different job with different tactics: comparison sites, communities, the specific review publications that turn up, and the small independent blogs that turn out to hold more slots than the market leaders do.

If vendors do hold slots in your category, as they do in help desk, accounting and payroll, then your own pages are a real route and the work is the ordinary one of making them the best answer to the question.

There is also a hopeful reading in the data that is easy to miss. The slots that do not belong to media giants mostly belong to very small sites. ifourtechnolab.com, marshadigital.com, myofficeops.com, steel-city.co.uk, customcpa.ca, ecomm.design. These are not household names, and several of them are holding stable citation slots that Mailchimp and Shopify are not. The slot is winnable by a small site. It is just not winnable by being the vendor.

Being citable at all is a precondition for any of this, and it is worth checking first, because a crawler that never reaches your pages produces the same silence as a page nobody wants to cite. It is also worth remembering that the exact wording matters enormously, since four phrasings of one question can share no reliably cited source, and that the durable slots in a category tend to belong to intermediaries rather than the organisations being asked about, which is the same finding this study now puts a number on.

Limits, and they are real

One surface. Everything here is Google’s AI Overview. ChatGPT, Gemini and AI Mode retrieve differently and would need their own measurement. Nothing here should be read as a claim about them.

One market for the headline figure. The 84 slots and the 24 are United States, English. 7 categories were re-run in Spain and the same 7 in Germany, and are reported above: enough to show that the ordering travels and the zeroes do not, and not enough to give those markets their own headline number. France is not unmeasured, and this sentence said it was until 16 August 2026: the register has held 6 comparable French cells since 12 August, measured the same way. They are not in the table above because the table is the three markets this study ran itself, and folding a fourth in without its own wording control is what the next pass is for.

Three runs. Three is enough to separate a stable source from a coincidence and not enough to estimate a rate precisely. A domain at two of three is genuinely uncertain, and this study calls it absent rather than guessing.

Eleven categories is a small sample of categories. The share of vendor-held slots ranges from zero to 83 per cent across them, which is a wide spread on a sample this size. The four zero categories are a real observation about those four categories. Extending it to “most categories” would need many more.

One question per category. Each category was probed with a single phrasing. Given how much phrasing moves these results, a different wording of the same intent could return a different set. The control question is the only place a second phrasing was tried, and it changed the answer completely, which is a caution as much as a finding.

Nothing here is causal. No intervention was run. This is what the answers contained on one day, not evidence that changing a page would change a slot.

The vendor split is a judgement. Ours is published domain by domain so that it can be checked and disagreed with.

Common Questions About Citation Slots

What is a durable citation slot?

A domain cited in every run of the same question. Asking once tells you what an engine reached for on that request. Asking repeatedly and keeping only the domains present every time separates the sources an engine relies on from the ones it happened to pick. With three runs the filter is strict, because a single absence removes a domain.

Why does the market leader not appear in its own category?

Because the question asks for a comparison, and a vendor’s own page is a poor source for a comparison of vendors. The control in this study shows the same vendor being cited reliably when the question is about it rather than about its category, so the absence is a property of the question rather than of the company.

Does this mean content marketing does not work for vendors?

No, it means it works in some categories and not others for this kind of question, and that you can measure which before you invest. In help desk software, five of six durable slots are vendor-held. In web hosting, none are. Same buyer, opposite answer.

How many runs do I need to do this myself?

Three is a workable minimum and it is what this study used. Two is not enough, because with two runs “present in every run” has no tolerance at all and any statistic built on it moves violently for mechanical reasons. If you only have budget for one pass, treat the result as a single observation and say so rather than reporting it as a rate.

Can I just repeat the same prompt in a chat window ten times?

Not usefully, if the tool you use serves from a cache. Repeated identical prompts can return byte-identical text, in which case ten readings are one reading with decoration. Whatever you use, verify that repeated runs actually differ before reporting anything with a denominator larger than one.

Does this transfer to ChatGPT or Perplexity?

Unknown from this data. Different engines retrieve differently, and this study measured one surface. Treating a result from one surface as a fact about “AI search” is the most common error in this field, and it is not one worth copying.

Where this leaves you

The question most companies ask is “how do we get cited”. The answer this data suggests is that for a whole class of high-intent questions, the honest first step is to find out whether a company like yours gets cited for that question at all.

That is a cheap thing to establish and an expensive thing to assume. Eleven questions, three runs each, one afternoon. Four of the ten categories with a stable core had no vendor in it, and any company in those four writing yet another page about why it is the best was writing for a slot that, on this surface and on this day, did not exist.