We Asked the Same Five Buying Questions in Six Markets. Whether a Vendor Gets Cited Tracks the Category, and One Market Does Not Follow.

Our cross-day study found that vendor domains held a durable citation in 3 of 19 United States slots and 7 of 9 Spanish ones, and said in print that the gap needed its own study before anyone called it a market difference. This is that study. Asking web hosting, ecommerce platforms, CRM, payroll and accounting software in the United States, Spain, Germany and France, on the same surface with three runs each, the pooled vendor share climbs cleanly by category: 4, 19, 22, 22 and 38 per cent. Per market it holds in only two of the six, and the sharpest result is elsewhere: both same-language pairs we have, the United States against the United Kingdom and Spain against Mexico, put the second market far below the first.

The market difference was mostly reading which categories each market happened to have.

Disclosure: EchoWi sells AI visibility measurement, so a study about which categories let a vendor hold a citation is a study we have an interest in, and it argues that a good deal of what buyers want to move is not movable by them. The four questions, the four markets, the surface, the run count and the dates are below, and all thirty rows with every domain named are in our measurement register, so anyone can rerun it against us.

The short version

1. By category, pooled across six markets: web hosting gives vendors 1 durable slot of 27, ecommerce platforms 4 of 21, CRM 6 of 27, payroll 9 of 41, accounting software 14 of 37.
2. By market, pooled across the five categories: the United States 33 per cent, Spain 36, Germany 17, France 15, the United Kingdom 5, Mexico 17.
3. Same language, second market, roughly a third of the share, twice. English: 33 per cent in the United States against 5 in the United Kingdom. Spanish: 36 in Spain against 17 in Mexico. One pair is an accident and two is a pattern worth naming.
4. The per-market ordering now holds in two of six, and the pooled gradient still ascends. Applying our classification rule uniformly, which we had not been doing, cost the ordering one market on its own.
5. The classification rule bites unevenly by market, and that is a confound for the market half. Spanish-language answers are full of suites that call themselves business management software, which a rule requiring the page to name the product leaves out. English answers are full of single-product vendors, which it keeps.

What a durable slot is, and why it is the thing worth counting

A durable slot is one domain that came back in every answered run of a buying question. Not cited once, which a single answer can produce by accident, and not present on the results page, which is a different surface. It is the set a category returns when you ask it three times and keep only what survives all three.

The reason to count those rather than mentions is that they are the positions a brand can actually be in. If a category has eight durable slots and none of them belongs to a company selling the product, then no amount of work on a vendor’s own site changes that answer, because the answer is not built out of vendor sites.

Each of the thirty rows here is one buying question, asked on one answer surface with three runs. Five categories, six markets, and every row answered all three times, which is what makes them comparable to each other. The Spanish payroll question needed two attempts to get three answered runs, and the second returned the identical six domains as the first.

The thirty rows

Read down the category blocks and the gradient is hard to miss. Hosting hands vendors one slot out of twenty across four countries. Accounting hands them fourteen out of twenty-seven, and the three categories in between land in order.

Where the gradient holds by market, and where it inverts

Two of the six markets rank the five categories in the same sequence, left to right, with no inversion: the United States and Germany. The other four each invert somewhere, and every inversion in the table sits between two cells carrying zero or one vendor slot. France does not: its payroll question returns seven durable domains and no vendor among them, which puts payroll below hosting and breaks both the sequence and the low end of it.

France is worth stating precisely rather than explaining away. Its five cells carry zero or one vendor slot each, so its ordering is decided almost entirely by the denominators, which run from three to eleven. A category with one slot of four outranks one with one slot of five, and nothing about those two categories caused that.

The other half of the French row is a classification limit rather than a measurement. Its payroll answer is held by an accountancy firm, an HR magazine, two workforce scheduling products that do not describe themselves as payroll, a software integrator and one domain that returns a bot challenge to every method we have. The two domains in that answer that do sell payroll each appeared in one run of three, so they are not durable. A different reading of the two scheduling products would move that cell from zero to two.

So the claim has narrowed twice as the study grew, and that is what an honest test looks like rather than a defect in it. Three categories gave an ordering identical in all four markets. A fourth broke it in France. A fifth broke France’s low end too. What survives is the pooled gradient and the observation that four of the six markets invert somewhere.

The French payroll zero was the wording, and that narrows this study’s own result

Measured on 16 August 2026, after the section above had been published in four languages.

The sentence above breaks France out of the ordering on one number: its payroll question returns seven durable domains and no vendor among them. Our own rule says that when a market falls outside a pattern, the first hypothesis is that the wording is different rather than the market, so the question was asked a second way. Quel est le meilleur logiciel de paie pour une PME ? became Quel logiciel de paie choisir pour une TPE ?, same surface, same market, three runs, and the second wording enters the register as a control rather than as a sixth French cell.

It returns 7 durable domains again, and one of them sells payroll: malibou.com, whose own home page calls it Logiciel RH & Paie. That is 1 vendor of 7, exactly what French web hosting returns, so under the second wording payroll does not sit below hosting and the low end of the French sequence is not broken.

Three domains were opened and rejected rather than counted. staffngo.com sells an HRIS and advertises that it integrates with payroll, which is the same reason the scheduling products were excluded from the first reading. tool-advisor.fr is a software comparison site. tplpaye.fr outsources payroll as a reseller of two other vendors, and reaches two runs of three in any case.

What this changes and what it does not. Not the table, which records what each wording returned on the day it was asked. Not the pooled gradient, which is the result this article leads with. What it changes is the confidence of one sentence: France failing to reproduce the ordering is now a statement about one wording of one question, and the other four French cells have not had the same treatment. Settling it either way needs a second wording for all five, which is the section below.

It is the second French zero to dissolve this way. The accounting question returned eleven durable domains and no vendor, and asked a second way it returned Pennylane. Two for two is not a rule. It is enough to say that a French zero in this study should not be quoted without its control.

All five asked a second way, and France stops inverting for the wrong reason

Measured the same day, three runs each on the same surface and market, with the prediction written down first: a market property would keep France’s ecommerce-above-CRM inversion, a wording property would remove it.

The inversion is gone and that is not a vindication. Read as shares the second wording gives 0, 0, 0, 14 and 40 per cent, which never decreases, so France passes the ordering test it failed. It passes because its bottom three cells all went to zero, not because anything reordered. A no-inversion test is easy to pass on ties, and three ties is most of this row.

What actually moved is larger than the ordering question. Vendor share fell in three cells and rose in two, and no cell landed within five points of where the first wording put it. The same market, the same day, the same surface and the same five products: asking Quel X choisir pour une TPE ? instead of Quel est le meilleur X pour une PME ? changes how many of the durable citations belong to vendors in every one of the five.

One classification call decides the headline, so both counts are here. monday.com holds a durable slot in the CRM answer and its home page sells an AI work platform without naming CRM, which is the rule that already excluded a French logiciel de gestion from accounting and an HRIS from payroll. Not counted, France reads 0, 0, 0, 14, 40 and does not invert. Counted, CRM becomes 1 of 5 and lands above payroll, which is a fresh inversion in a different place. The claim that France does not reproduce the ordering survives under one of those two readings and not the other, and we are not going to pretend the rule picks itself.

So the honest state of this row is narrower than either version we have published. France’s ordering is decided by cells carrying zero, one or two vendor slots out of three to seven durable domains, and it moves when the question is reworded. That is a statement about how little a single market cell carries, and it is the reason the article leads with the pooled gradient instead.

The same language twice, and the second market is far below the first

This is the result that came out of adding markets rather than categories, and it is the only one here that has replicated.

Two pairs, same five questions, same strings inside each pair, and in both the second market returns between a seventh and a half of the first market’s vendor share. One pair is a coincidence. Two, in the same direction, is worth naming and worth attacking next.

What it is not, yet, is explained. The obvious candidate is that these answers lean on national publishers, and a smaller national publishing layer leaves fewer slots of any kind, but this study counts vendor slots as a share of durable slots and so already divides that out. The other candidate is our own classification rule, and the next section is about why that one cannot be dismissed.

The classification rule bites unevenly, and that is a confound for the market half

Our rule is that a domain counts as a vendor when its own home page names the product the question asks about. Applied uniformly for the first time in this version, it removed four calls we had made loosely: a Spanish business-management suite in CRM, a Spanish HR product and a Spanish business-software suite in payroll, and a British all-in-one finances product in accounting.

The cost is visible and one-directional. Spanish-language answers are full of suites branded as software de gestión empresarial, whose pages describe a suite rather than a product; the rule leaves them out. English-language answers are full of vendors that name one product on the front page; the rule keeps them. Mexico’s payroll cell is zero of five under the rule and three of five if the two suites count, and its accounting cell is zero of four or two of four the same way.

So the market comparison carries a rule that does not bite equally across markets, and the category comparison does not, because every category is measured in every market with the same rule. That is why this piece leads with the gradient and treats the market numbers as the part that still needs a better instrument.

The alternative rule, counting a suite when it plainly contains the product, needs us to know each vendor’s product line, which is exactly the kind of judgement that made the first version inconsistent. Publishing both counts wherever they differ is the compromise, and every affected cell carries its other number in the register.

Why the earlier gap looked like a market difference

The cross-day study compared whatever categories each market had. The United States had eleven, Spain five, and they were not the same eleven and five. When one market’s set leans toward the categories that cite vendors, its vendor share rises without anything about the market changing.

Matching the categories shrinks the market spread but does not erase it: 33, 36, 17, 15, 5 and 17 per cent on the same five questions. Spain is still the highest, and the reason is visible in the table above. Nine of Spain’s twelve vendor slots are in payroll and accounting, and Spanish accounting is the single densest vendor cell in the study at five of seven.

So the honest answer to the open question is a partial one. Category mix explained most of the original gap. What remains is a difference inside accounting, where the four markets run 33 to 71 per cent, and that sits on three to thirteen slots per market, which is a recount and not a rate.

What occupies the slots when vendors do not

The same handful of general purpose domains, again. Counting how many of the thirty rows each domain held a durable slot in: YouTube in twenty-four, Reddit in seven, and a German founders’ portal in five. Nothing else appears in more than two.

Everything else is national. French hosting is held by a hardware and software magazine, a developer forum, a digital agency and an SEO agency. German ecommerce is held by a marketing media brand, a founders’ portal and a marketing agency. Spanish accounting is the exception that proves the shape, because it is where the vendors are.

Two classifications are worth naming because the domain name would have got them wrong, and getting them wrong would have moved a figure in this article. apogea.fr holds a durable slot in both French accounting and French help desk, and it sells integration of management software rather than the software, so it is not a vendor in either. payplug.com holds one in French ecommerce and sells payments, which is adjacent to a store platform rather than being one. Both were settled by opening the home page, which is the only method that works: a domain name is a hypothesis about what a company sells.

What this means if you sell in one of these categories

If you sell hosting or an ecommerce platform, the answer is not made of vendor sites. One durable slot in twenty across four countries is not a target to optimise toward, it is a description of who gets to be in the answer. The work that pays is placement in the sources that do hold the slots, which here are magazines, forums, agencies and video. We counted that layer afterwards and it changed this advice: 79 per cent of the domains holding it appear in exactly one cell, so video is the instruction that survives and the rest is a list of coincidences.

If you sell accounting software, the opposite. Fourteen of twenty-seven slots go to companies selling the product, so a vendor site is a legitimate candidate and the usual work on answerability applies to it.

And check your own category before assuming which of those you are in. The two cases are a factor of ten apart, and nothing about the question wording or the market tells you which one you have. Three runs of one question is enough to find out.

What this does not show

Five categories and thirty rows. This is a small study by design, because matching categories across four markets costs a row in every market for every category added. It is enough to establish a pooled gradient and to find one market that does not follow it, and not enough to put a number on any single cell.

One surface, one reading each. Every row is a single reading on one answer surface. Our own cross-day work shows a second reading of the same row moves it, so the individual cells here should be read as one draw, and the pattern as the finding.

Small numbers, stated as counts. French accounting is one vendor slot of three. German accounting is two of four. Those are recounts, and they are printed as recounts rather than as 33 and 50 per cent for exactly that reason.

The market question stays open. Matching categories shrank the gap and did not close it, and the residual sits inside one category on very few slots. Settling it needs the same five questions in more markets, or more categories in these six, and a classification rule that does not depend on how a vendor brands itself.

Nothing here is causal. No intervention was made. This describes what a surface returned on one day, not what would happen if a vendor changed anything.

Vendor classification is a judgement. Ours is published domain by domain in the register so a reader can disagree with a specific call rather than with the total. Several rows turn on a call we made by reading a home page, and the ones that would move a figure are named above. Two of them, ahora.es and santanderx.com, return nothing at all to a plain fetch and were read in a browser, which is the same rule our price work follows.

Common Questions About Vendor Citations by Category

Do AI answers cite vendors or publishers?

Mostly on the category. Across five buying questions asked in four markets, web hosting gave vendor domains 1 durable citation slot out of 27 and accounting software gave them 14 out of 37, with ecommerce, CRM and payroll landing in between at 19, 22 and 22 per cent. That pooled gradient still ascends, and per market it reproduces in only two of the six.

Why does my category never cite my website?

Because in some categories the answer is not assembled from vendor sites at all. Web hosting and ecommerce platforms returned almost no vendor slots in any of the four markets, and the domains holding those slots were magazines, forums, agencies and video. That is a property of how the category is answered rather than a defect in a particular site.

Does the country change which sources get cited?

The specific domains change completely, and the shape mostly does not. Every market in this study returned its own national publishers, with only YouTube, Reddit and one German founders’ portal crossing between them. Which categories give vendors durable slots ranked the same way in two markets of six, and the level changes a great deal: two pairs sharing a language put the second market at a seventh and a half of the first.

Can I use a vendor share measured in one market for another?

Use the ordering rather than the number, and check that it holds where you are. And check it on the surface you care about, because the ordering does not transfer between surfaces either: the same five questions on AI Mode put a different category at the top. The pooled market share moved between 5 and 36 per cent across six markets on identical questions. The ranking of categories held in two of those six, and the two same-language pairs in the set each put the second market far below the first, so a rate is local and a ranking is a hypothesis worth testing rather than a constant.

How do you decide whether a domain is a vendor?

By opening its home page, never by reading its name. In this study a French domain holding two durable slots sells integration of management software rather than the software itself, and another sells payments rather than store platforms. Both would have been counted as vendors from the domain name alone, and both would have moved a published figure.

How many runs does this kind of measurement need?

Three at minimum, and the durable set has no tolerance at three: a domain missed by one run drops out entirely. Every row in the comparison answered all three times, which is stated per row in the register. One question needed a second attempt to get three answered runs, and both attempts are recorded rather than only the one that worked.