We Asked an Assistant for Alternatives to Eight Vendors. One Answer Was About a Different Company Entirely.

We asked an assistant “what are the best alternatives to Conductor?” and it recommended, among other things, an open-source manager for running coding agents in parallel. That is a real product. It is not the SEO platform anyone asking that question meant. Two of the ten sources it cited were about the wrong company, because two companies share a name and nothing in the question told the assistant which one you meant. We asked the same question about eight vendors. The homonym was not even the strangest result. In 6 of the 8 answers, the vendor you asked about was not cited at all.

Disclosure and method: EchoWi sells AI visibility measurement and every vendor here is a competitor. So the method is mechanical: one question per vendor, AI Overview throughout, every domain classified by opening its homepage and never by reading its name. The main arm is 8 vendors in the United States in English, 21 August 2026, 5 runs requested and between 2 and 5 answered. The fourth market below is the same 4 vendors in Germany in German, the same day, 6 runs requested and between 4 and 5 answered. Two further arms are the controls below, all 8 vendors in the United States asked twice without the word alternatives. In each case the design, the deciding statistic, the prediction and the retirement threshold were written down and committed before the first call. All 20 alternatives questions and the 29 control questions are in our measurement register.

The short version

1. Ask for alternatives to a vendor, and that vendor usually is not in the answer. Its own domain was cited in 2 of the 8 answers.
2. One answer was about a different company. “Conductor” is both an SEO platform and a developer tool, and the assistant blended them.
3. 62 distinct domains were cited across the 8 questions. 20 of them are not in this category at all.
4. Whether this layer is self-serving depends on how you define a vendor, and the two reasonable definitions land on opposite sides: 34 per cent against our own catalogue, 61 per cent against what the homepages say.
5. That gap is our own measurement error, published rather than quietly fixed. It is also the most useful thing here, and the last section explains why.
6. There is no European layer, and the fourth market makes that sharper rather than weaker. In 4 of 4 vendors the German answer sits closer to the American one than to the French one, and the most distant pair in the whole table is two European neighbours.
7. We ran the control that could have killed this, and it holds. Asked a question that names the vendor without asking for alternatives, all 8 of 8 cite their own domain, against 2 of 8 when the question asks for alternatives.

The answer was about a different company

The question was “what are the best alternatives to Conductor?” and nothing about it is ambiguous to a human in marketing. Conductor is a well-known SEO and content platform.

To an assistant with no context, the name is shared. Of the ten domains cited in that answer, two are unmistakably about the other one: a visual editor for coding agents, and an open-source manager for running them in parallel. The rest of the answer is a mix, with a research firm and a jobs-and-reviews site sitting next to companies from neither category.

Nobody wrote a bad page here. The vendor did not make a mistake. The question simply does not contain enough to disambiguate, and the assistant answered both readings at once.

This is the failure mode nothing on your website fixes. You can write the best comparison page in your category and it will not stop an assistant from answering about a company that merely shares your name. What it does change is what you should measure. If your name collides with another product, the useful number is not your position. It is what fraction of the answer is even about you.

Ask for alternatives to a vendor, and the vendor vanishes

This was the prediction we wrote down before measuring, and it held. Across the 8 questions, the named vendor’s own domain appeared in 2 of the 8 answers.

It makes sense once you say it out loud. “Alternatives to X” is a question whose whole purpose is to get away from X, so an answer that leads with X has misread it. But it has a consequence worth sitting with if you sell software: the page you wrote about yourself is not what answers the most commercially loaded question anybody asks about you. Somebody else’s page is.

So who fills the space

62 distinct domains across 8 questions. Opening every homepage rather than guessing from the name, they fall into four groups.

Two things stand out.

A third of what gets cited is not the category. Twenty domains are review sites, agencies, a development shop, a link-building service, and the two coding tools from the Conductor answer. Somebody searching for a way to measure their AI visibility is being handed a mix in which one in three sources sells something else.

Classification by name would have been wrong repeatedly. One domain reads like a competitive-intelligence tool and is one; another has a name that sounds like a consultancy and turns out to sell exactly this software. The homepage is the data and the name is a hypothesis, which is why the four buckets exist rather than two.

Our own rule had a flaw, and fixing it flips the answer

Here is the part we would rather not publish, and the reason it is the most useful section.

The prediction written before measuring was that most cited domains would be published by a vendor in this category. That mirrors the ordinary search results, where 15 of 18 organic listings for these questions are written by a competitor placing itself first.

To test that, the frozen rule defined “a vendor” as one in our own verified catalogue. That is a clean, checkable definition. It is also the wrong one, and the size of the error is visible:

The two land on opposite sides of the threshold the prediction named. The reason is not subtle. 17 of the 62 cited domains are companies selling exactly this software, and our catalogue does not contain them. A rule that counts vendors by looking them up in your own list measures how complete your list is, not what the market looks like.

Both numbers are published, labelled, because that is what the evidence supports and choosing between them after seeing them is how a study gets an answer it wanted.

The transferable version, for anybody running this on their own category: if you classify cited sources against a list you maintain, your share of “competitors” is capped by your list. Ours was missing about a quarter of the field, and the field is young enough that anyone’s list is.

The fourth market, and Germany looks more like the United States than like France

This study had three markets and one sentence resting on them: there is no European layer. That sentence counts European markets, and it had two. Measuring France again would have measured France better. A third European market is the only thing that can knock it down, so we asked the same four vendors in German in Germany on 21 August 2026, against a design frozen and committed before the first call.

It does not knock it down. It sharpens it, in the direction nobody predicted:

The highest pair in the table crosses the Atlantic and the lowest sits between two European neighbours. Germany shares three times as much with the United States as it shares with Spain.

And the claim that needs no median over four cells is the one this section rests on: in 4 of 4 vendors the German answer is closer to the American one than to the French one. Vendor by vendor, without exception. Anyone buying Europe as one market is buying an assumption this table contradicts in every row.

The prediction was ours and it is refuted. Written down before the first call: if Europe were a layer, Germany against Spain and Germany against France would both exceed 0.30 while Germany against the United States stayed below 0.15, and we would have rewritten the sentence in four languages. The opposite happened, and it is published because a prediction adjusted after the measurement was never a prediction.

Across all four markets the answers cite 113 distinct domains, and 6 of them appear in all four: G2, Gartner, Reddit, YouTube and two vendors of the category. They are exactly the same six as with three markets. A fourth market added nothing to the intersection and took nothing away.

The homonym for the fourth time, and this time it refines the finding

In English “Conductor” collides with a tool that launches coding agents in parallel. In Spanish it collides with a common word as well, and the answer cited a windscreen chain’s blog. In French it collides with a multi-agent IDE.

In German it is the same family as English, only harder. The two domains cited in every answered run are agentsroom.dev, a terminal for multi-agent orchestration, and runpane.com, an open-source manager for running Claude Code, Codex and Aider side by side. Five more domains in that cell belong to the same cluster: 7 of the 20 cited domains are about the wrong company.

That is the more interesting reading, and it corrects what we would otherwise have written. It is not that every language meets its own separate product. It is the market that decides which meaning your name collides with, and two markets can catch the same one.

One exception belongs here because it runs against our own headline: conductor.com is cited in the German cell. No other market manages that. It happens in 1 of 4 runs and the domain sits nineteenth of twenty.

What this changes if you sell in this category

• Your alternatives page is not the main event. The answer to “alternatives to you” is mostly other people’s pages. The useful question is which of those pages, not whether yours ranks.
• If your name is shared, measure how much of the answer is about you. Position is meaningless when a third of the citations describe another company.
• A competitor list you maintain by hand goes stale faster than the category moves. 17 companies were cited here that ours did not contain, and it is refreshed constantly.

You can run this yourself, and the procedure is the whole method. Ask the question more than once. Write down every cited domain with how often it appeared, and open each homepage instead of judging by the name. It takes a few runs per question. If you would rather keep it running across surfaces and over time, that is what we build.

We ran the control that could have killed this study, twice

Everything above varies the context: another market, another language, another wording. None of it varies the thing the headline names, which is the word “alternatives”. If asking for alternatives is what pushes a vendor out of its own answer, then asking about that vendor without the word should put it back.

So we asked all eight vendors, same market, same surface, same day, twice. The controls started on four and were widened to the other four afterwards, because the claim they produce counts vendors and four is not a rate.

The first control was inconclusive on its first four cells, and it was inconclusive for a reason worth more than the number. “Is X worth using?” drops the word alternatives, and it drops the thing that told the assistant what X is along with it. It moves two variables. So we wrote down that a cleaner control was owed, and ran it: a question that still does not ask for alternatives but pins down the category without naming any vendor’s own.

Eight of eight. Every vendor is cited in its own answer once the question stops asking for alternatives to it. Predicted 7 or 8 before the first call of the widened run, and the threshold that would have sent the claim back to the original four was 5 or fewer.

And the middle row is not noise between the two: it is the point. The three steps are 2, 4 and 8 of the same 8 vendors on the same day, and we first read that as a gradient in how much the question pins down what you are. A fourth form, measured a day later, says that reading was wrong. It is two sections below, and it is the form buyers actually send.

So the headline holds, and now it holds for a stated reason rather than by default: the alternatives framing is what keeps a vendor out of its own answer, not some general property of how these answers get built.

One thing came free and is recorded rather than claimed, because we did not predict it: the clean question answered 2 to 5 runs per cell against 1 to 6 for the vague one.

The page that wins a five-vendor comparison is none of the five

Every form above names one vendor. The form the market actually sends names five: a head vendor, four competitors, one aspect. It arrives as a comparison request often enough to be the most common shape we see, so we ran it.

Same eight vendors, same four companions in every cell so that the only thing changing is the head name, AI Overview, United States, English, one day, six runs requested. The design, the deciding statistic and a retraction threshold went into writing before the first call.

One domain is cited in all eight answers, in every single run that answered, and it is not one of the five names in any of the questions. Rankability sells AI visibility for agencies. It was never asked about. It is in every answer.

The fourth row is the one to sit with if you sell here. Profound’s own site is cited in 5 of the 8 cells, and 4 of those 5 are comparisons about somebody else. One vendor is better represented in its rivals’ comparisons than most vendors manage in their own.

And the gradient we published a day earlier does not survive it

Here is the part that costs us something. A five-vendor comparison pins the category down harder than any question above: it names five products in it. If the gradient were about specificity, this form should have been the top of it.

3 of 8. That is the number, and here is what it costs us.

The threshold for retracting the specificity sentence was 4 or fewer, written down before the run. Three is under it, and it holds if you throw away the one cell that only answered once: 2 of 7.

So the sentence comes down and a better one replaces it, because all four rows fit a single rule. Your own site is cited when the question is about you alone. Ask for alternatives and you are excluded by the request. Name four rivals and you are one fifth of the subject. The only form that puts a vendor’s own page in front of a buyer is the one that asks about that vendor and nothing else.

That is worse news than the version it replaces, and more useful. Specificity is something you can write into your pages. How many competitors a buyer types is not.

One prediction did hold, and it was written down too: the homonym stays dead. The five-vendor form returned zero cosmetic-medicine domains for Profound and zero orchestras, hi-fi magazines or code orchestrators for Conductor. Naming four rivals in the category disambiguates as completely as naming the category did.

We tried to break that finding, and it held for a better reason than we thought

The eight questions above all named the same four companions. That was on purpose, so the only thing changing between cells was the head vendor. It also means the finding had a hole in it, and we only saw the hole by asking which Rankability page was being cited so we could write about it.

The answer came back hubspot-aeo-alternatives and hubspot-aeo-review. HubSpot AEO was in all eight questions. So two readings fit the same data: the domain owns the comparison layer, or it owns the pages for the one name we repeated eight times.

We ran the same shape with that companion swapped out. 3 of 4 answered cells still cite it, with HubSpot AEO nowhere in the question. The threshold was written down first, and 0 or 1 would have sent the sentence back for a rewrite.

And the page it wins with is not what we predicted. We expected a page named after whichever vendor the question mentions, the way the HubSpot ones were. What gets cited instead is how-much-should-you-pay-for-ai-search-visibility-tracking-tools, which names nobody. Its first line is a question a buyer actually asks and a priced answer you could quote whole: budget twenty to ninety-nine dollars a month, start here.

So the domain does not win by having one page per rival. It wins by having both shapes at once, so that whichever way the question lands there is a page waiting. The per-vendor pages catch the questions that name a vendor. The category question catches the rest.

That is the most copyable thing in this study, and it costs nothing to act on: the page that beat five named vendors to the citation is a self-contained answer to “what should this cost”, with the number in it.

One instrument note, because it changed the design mid-run. The replacement companion we had written down was Scrunch AI, and that variant returned no AI Overview at all, four runs and four failures across three head vendors, while the original question answered on its first run in the same window. That is a fact about the surface, not about anybody’s visibility, so the companion was swapped again and the failed variant is in the register with zero answered runs.

The page that wins every American comparison does not exist in Germany

Rankability is in all eight American answers. We ran the same shape in German, in Germany, with the same four companions, on the same day.

Rankability: 0 of 3 answered cells.

That number was the threshold. Written down before the run: 0 or 1 of 4 means the comparison layer localises like everything else this corpus has measured, and 3 or 4 would have made it the first layer we have seen travel between markets. It localises.

What fills the space is local instead. All 3 German cells cite German domains that appear in none of the eight American ones: a software comparison site, a German marketing publication, HubSpot’s German subdomain.

The fourth cell is not a zero and is not counted as one. It timed out five times while its three siblings answered the same day with the same shape, so it is recorded as unanswered and excluded from every count. A timeout read as an absence is a fact about the instrument wearing the clothes of a fact about the answer.

And one thing did travel: the finding itself. 2 of the 3 German cells cite the named vendor’s own domain, against 3 of 8 in the United States. The form behaves the same way in both markets. The sources behind it share nothing.

That is the practical shape of it. If you sell in more than one country, the page that owns the comparison in one of them buys you nothing in the next, and there is no reason to think the winner is the same company twice.

The same name, three different companies, depending on the question

Here is what the controls found that nobody was looking for.

Ask Is Profound worth using? and the answer is not entirely about software. Profound is also a radiofrequency skin-tightening treatment, so the sources come back split: a college of medicine, four med spas and a plastic surgery practice sit in the same answer as the AI visibility vendors. Of the 27 sources it cited, 6 were confirmed to be cosmetic medicine by opening them, and 6 more answered a bot challenge and are not counted, so 6 is a floor.

Now add five words. Is Profound a good tool for SEO teams? cites 10 domains and not one of them is cosmetic medicine. Conductor does the same thing: the open-source coding-agent cluster that took every single run of its German alternatives answer is simply gone, and conductor.com is cited in 5 of 5.

That was a prediction too, written before the calls with its threshold: fewer than 2 confirmed cosmetic domains. It came back zero.

And it replicates in a second market. The same four vendors, the same three questions, in German in Germany on the same day and surface, give 1 of 4, 2 of 4 and 4 of 4. Germany was chosen because it is the only European market whose alternatives baseline is not zero, so the result could not be a floor effect. The German collision is a different one and better: asked Lohnt sich Conductor?, the domain cited in every single run is the personal site of an orchestral conductor, followed by four hi-fi magazines, because Conductor is also a headphone amplifier. 6 of the 15 domains were confirmed to be about something else by opening them. Ask the same company with five more words and none of them is left.

What this means if your brand name is also something else. Your visibility is not a property of your brand. It is a property of the question, and by a lot: the same name, the same day, the same surface, and three questions produce an answer about a software vendor, an answer about a skin clinic, and an answer that never mentions you at all. A tool that runs one phrasing per brand and hands you a percentage is measuring one of those worlds and calling it yours. Running several phrasings per intent is the whole reason our own measurement exists, and it is what we build.

What this does not show

20 alternatives questions and 29 control questions, one surface, 4 markets, 4 languages, one day. AI Overview only. Its citation list returns every marker it prints, while other surfaces truncate or hide the publisher behind a redirector. Nothing here describes ChatGPT, Gemini or AI Mode.

Between 2 and 5 runs answered per question in the main arm, and between 4 and 5 in the German one, so the size of a stable set is not comparable between them. That is why the deciding statistic is a share of all cited domains, which does not move with the number of runs. The market table carries the same caveat in a sharper form: two Spanish cells answered exactly once, and a small set mechanically depresses any Jaccard against a large one. That is why the set sizes sit beside the figures, and why the section rests on the vendor-by-vendor comparison rather than on the median.

4 homepages returned a bot challenge and were not read. All 4 are counted as outside the category, which lowers our own category counts rather than flattering them, and they are flagged as such in the register.

A correction to ourselves: Profound’s own domain was recorded wrongly in three already published arms. It is tryprofound.com. It changes no figure, because neither spelling was ever cited, and that is exactly why it is written here rather than quietly fixed.

A citation is not a recommendation and not a click. We measured which domains the answer cited, not whether anyone read them or what the answer said about them.

Frequently asked questions

Is it worth writing an alternatives page at all?

It is worth writing, and not for the reason usually given. In ordinary search results these questions are dominated by vendors publishing lists that place themselves first, so the page competes in a crowded and openly self-interested field. What our measurement adds is that the assistant’s answer is drawn from many such pages at once, so the realistic goal is being one of the sources it draws on, not being the answer.

My company name is also a common word. What do I do?

Measure the contamination before anything else. Ask the assistant the questions your buyers ask, then check what fraction of the cited sources are about your company rather than the other one. If a meaningful share is about somebody else, your visibility number is describing two companies and cannot be improved by editing your site. Context in the question fixes it for a human asking; nothing on your side does.

Why only AI Overview, and not ChatGPT or AI Mode?

Because this study asks which domain published a cited page, and only AI Overview answers that reliably by this route. AI Mode wraps the publisher in a redirector so the domain field reads the same for every citation, and it returns only some of the citations its own answer prints. Using either would have produced a cleaner-looking table built on a field that does not mean what it appears to.

How many runs is enough?

More than one, always, and the reason is visible in this data: between two and five runs answered per question, and the set of cited domains changed between runs in every cell. A single run tells you what one answer said, which is a photograph and not a rate. Anything published as a percentage needs repeated runs that skip the cache, and anything published from one run should say so.

Does a cited domain mean the assistant recommends it?

No, and conflating the two is the most common error in this area. A citation is a source the answer drew on. The company being recommended in the text and the domain being cited underneath are separate outcomes, and in our measurements they come apart often enough that any serious tracking has to record them as two different things.