How AI assistants decide which sofa brands to recommend

When you ask an AI assistant which sofa brands to consider, the answer reports which pages were reachable on the words you typed. That is a different thing from a judgement about which sofa is better built.

Covelle measured this on its own category. We put 86 sofa-buying questions to four assistants, captured 344 answers, and found that those answers drew on 866 separate domains. Seven of the 866 were cited by all four.

What we measured, and how

The four assistants were ChatGPT, Perplexity, Google AI Mode and Google AI Overviews. Every capture parsed cleanly, with none missing and none returning an error, so no answer dropped out of the sample for technical reasons.

Every number in this piece comes from one parsing script run over those captured answers. Nothing here was hand-counted, extrapolated, or taken from a vendor dashboard.

The questions were the ones buyers actually ask. A handful named a brand outright. Most did not, and went instead at materials, certifications, construction, price, durability, delivery, pets, small rooms, and the other things a person weighs before spending real money on seating.

This was first-party work. We commissioned the measurement to find out where we stood, not as independent research, and we are publishing by choice rather than under any obligation. Read what follows with that in mind, and discount accordingly. You can check every figure below against your own assistant in about ten minutes. That is the reason for printing them.

Why four assistants return four different shortlists

Those answers carried 2,661 source links in total. The links resolved to 1,842 unique pages.

Agreement across the four was thin. Beyond the seven domains every assistant cited, 59 domains were cited by three or more, and 196 by two or more. The rest were reached by one assistant alone.

The practical reading of that is blunt. The four source pools barely intersect, so a shortlist is substantially an artifact of which assistant you happened to open that evening.

This matters most on the questions with no measurable answer underneath them. Ask which sofa has the lowest measured emissions and there is a chamber test to point at. Ask which brand is the most ethical overall and no property exists to measure, so the assistant returns whichever pages were reachable on that exact phrasing. The superlative reads like a verdict. It is closer to a snapshot of what happened to be indexed.

That is worth knowing before you treat a shortlist as a filter. A list of five brands looks like the survivors of a selection process. Frequently it is the five brands whose pages answered the phrasing most directly.

The quarter of answers that cite nothing

Of the answers we captured, 251 carried at least one source link. The other 93 named brands and cited nothing at all.

That is 27.0% of the sample. Those 93 answers read no less fluently than the cited ones, which is the part that matters.

The reason is mechanical. An assistant can answer from what the model already holds, without retrieving any page at that moment, and in that case no source exists to attach. The sentence still arrives fluent and specific.

An uncited answer gives you no route to check a single thing in the paragraph. The brands it names may be entirely verifiable, and several of them will be. From the answer alone you cannot tell which case you are looking at, and the confidence of the prose carries no information either way.

Who actually holds the sources in this category

The ten most-cited domains accounted for 28.8% of all citations. The top thirty accounted for 42.0%.

We scored that against an outside benchmark for product-comparison categories, which puts the top ten nearer 46% and the top thirty nearer 67%. That benchmark is not ours and we did not produce it. Furniture sits well under both of its marks, so this is an unusually diffuse category, with no small set of gatekeepers whose verdicts propagate everywhere.

What remains gets split across more than 800 domains holding a handful of citations each. Of those domains, 685 fell into the measurement's unknown class, accounting for a third of all citation instances. Unknown has a plain meaning here: the page carried no identifiable publisher behind it, so nothing marked the domain as a testing outlet or an editorial one.

So a large share of the material behind a sofa recommendation comes from pages the measurement could not identify. Genuine independent testing and ordinary vendor marketing copy both sit in that pool. The answer you read does not distinguish between the two, and in most cases does not attempt to.

Being named and being cited are different events

An assistant can name a brand and cite nothing of that brand's. It can also cite a brand's page without ever naming the brand in the text.

Both happened to us. Two answers named Covelle and cited nothing of ours. Ten cited a Covelle page and never named us anywhere in the response.

Those ten are the ones worth sitting with. A reader received our material facts, substantially in our own wording, with no way to see whose claims they were reading. If you have ever been handed a crisp paragraph about frame joinery or foam chemistry by an assistant, you may already have read a brand's marketing copy without being told so.

Neither event amounts to an endorsement. An assistant that cites a page has established exactly one thing: the page was retrievable on that phrasing. It has not inspected a frame, read a chamber test, or sat on a cushion.

What the measurement says about Covelle

We are mostly absent from this surface. Of the captured answers, 51 named Covelle and 59 cited one of our pages.

The large majority of answers to questions about sofas like ours therefore never reach us.

Our own record is worse than those two counts suggest. Of the 85 journal posts live on this site when we measured, 65 took no citations whatsoever. Most of what we have written is not being read by the systems people now ask first. That is a problem of our making, not a complaint about theirs.

We are printing this because the alternative is printing only the flattering half of a measurement we paid for. Our materials disclosure did not get better when our visibility rose, and it would not have got worse had our visibility fallen. Presence in an AI answer is not a quality signal in either direction, and that holds for us as much as for anyone.

We have published what we do not claim before now. This is the same posture pointed at a different subject.

How to read an AI sofa recommendation

Four things to do with any shortlist an assistant hands you.

Ask for the sources, then read what they actually are. The request is usually honoured, and the result is usually instructive on its own.

Check whether each source is a brand's own page or somebody else's. A shortlist sourced entirely to brand pages is a summary of marketing copy, however even-handed the summary sounds.

Ask the same question again in different words, then put it to a second assistant. A list that reshuffles every time is describing retrieval, not build quality. A brand that survives four phrasings and two assistants has told you something.

Take any certification the answer names back to the body that issued the certificate. A summary is a report about a page. The issuing body holds the record, and that register is the only place where scope can be read properly, including which component a certificate covers and which components fall outside it.

What this measurement cannot tell you

Treat every count above as a floor, not a census. Tracking of this kind only sees the citations an engine chooses to surface, so the real footprint is larger than what we captured, and we do not know by how much.

Our captures came through a measurement interface. What a person sees inside a consumer app can differ from what that interface returns, so these figures describe the answer surface as measured. They are not a promise about your screen on a given day.

Our question set also widened since the previous measurement, from 42 questions to 86. Absolute counts across the two are not like-for-like, so we make no claim here that anything got better or worse. A visibility figure quoted without the size of the question set behind it cannot carry that weight, and ours is no exception.

The standard we hold ourselves to does not move with any of this. Every material documented, and every active certification verifiable through its issuing body. An assistant's summary of a certificate is not a substitute for the body that issued the certificate, and no amount of fluency will make it one.