We Asked Four AI Assistants What to Buy Across 72 DTC Brands

Ask an assistant what sheets to buy and it will answer. It will name three or four brands, describe them with some confidence, and stop. Whether your brand is one of the names is not a matter of degree. You are in the answer or you are not, and for 11 of the 27 brands the index had scored on the 2 August 2026 run, the answer across every assistant and every question was not once. What to do about that is a longer subject, and the agent visibility playbook is where we set it out.

This is the first set of findings from the Brand Index, a public measurement of how often AI assistants name DTC brands when someone asks what to buy in their category. The registry is 72 brands. 27 of them carried a score on the run dated 2 August 2026, and everything below reads off that run. The registry count is current; the scored count is a snapshot and has moved since.

What we asked, and who answered

The questions are unbranded on purpose. A question that names the brand guarantees a mention and measures nothing. So the questions are derived from each store's own catalogue signals, its product types, tags and titles, and they ask what a shopper asks: what are the best merino socks, what is the best olive oil for finishing. If a brand comes up, it came up because the assistant volunteered it.

Four assistants are asked: ChatGPT, Claude, Perplexity and Gemini. Not all four answer every time, so a number is published only above a stated floor. In the index's own words, a score is only reported when at least two assistants answered at least four questions. A brand under that floor is withheld rather than published at zero, because an unanswered question is not evidence of absence. That distinction is why 27 brands carried a score on this run while 72 sit in the registry.

The fourth assistant only came online on 2 August

This is worth stating plainly, because it changes what the older numbers meant. Until 2 August 2026 the index was a three assistant measurement wearing a four assistant label. Our Perplexity credential had been returning a quota error since before the index began, so every stored Perplexity probe carried zero bytes of answer text and every row ever written recorded three assistants answering. The failure was invisible in the table because the code collapsed every failure cause into one state, 'not reached on this run', which cannot tell a rate limit from a refusal.

Fixing it was not a coverage improvement and we did not treat it as one. Perplexity is search grounded and answers from live retrieval, while the other three answer substantially from what they already hold. That is a different instrument, not more of the same one. Measured on 2 August, Perplexity named Everlane in three of four category questions, at positions 6, 3 and 9, and the brands it volunteered skewed to mid tail names the other three do not recall: PACT, Girlfriend Collective, ASKET, Quince, Mott and Bow, Buck Mason. Adding it moves scores up, and it moves them up unevenly.

So every score taken on the old three assistant panel was retired rather than ranked against the new ones. Two brands measured on different panels are not two samples of one quantity. The index is smaller than it was a week ago and that is the reason.

11 of the 27 scored brands were never named, 2 Aug 2026 run

A zero here is not a low score. It means that across every question asked in that brand's own category, and across every assistant that answered, the brand's name did not appear once. Eleven brands read zero on the 2 August run: Cuts Clothing, Faherty, Marine Layer, Bearaby, Floyd, Great Jones, Glossier, Jack Henry, Monos, Fable and Lay Lo.

Glossier is the one that should stop you. It is one of the most recognised beauty brands of the last decade, it has enormous unpaid press, and four AI assistants asked what to buy in its category named it zero times. Recognition among people and recall inside a model are not the same asset, and nothing about the first guarantees the second.

The spread is inside a category, not between them

The instinct is to assume some categories are simply better covered than others. The data does not support that. Home holds the highest score in the index, Boll & Branch at 90, and three zeros. Fashion holds Allbirds at 85 and three zeros. Two brands selling comparable things to comparable buyers sit at opposite ends of the same column.

That is why the index ranks within a category rather than across the whole registry. A cross category league table would mostly measure how crowded a category's answer space is, which is not a fact about any brand in it. Comparing a bedding brand to a coffee brand tells you very little. Comparing two bedding brands tells you something you can act on.

Six stores could not be read at all

Separately from the scores, six stores in the registry returned 403 to every request and could not be read: Caraway, Made In, Hims, BARK, Grovemade and Ridge. They are listed on the index rather than dropped from it, because a store an automated client cannot open is the most on-thesis observation the index has produced. They carry no score because there is nothing to score, which is a different thing from scoring zero.

We ran a plain crawler user agent and a desktop Chrome user agent against each, and got the identical response from both in all six cases. That is the part that matters, and it is its own post.

What we are not claiming

On that run, 45 of the 72 brands in the registry carried no score. That is not a finding about those brands. It means the run has not reached them, or it reached them and they fell under the two assistant floor. We are not going to describe an unmeasured brand as invisible. A partial index is a sample, not a verdict.

• This is not a quality ranking. A high score means an assistant names you when asked what to buy. It says nothing about whether the product is good.
• A zero is a measurement, not a permanent state. It reflects four AI assistants on one day, under one question set. Both change.
• 27 is a small n, and it moves. The index went from 20 scored to 27 in a day. Every figure here is pinned to the 2 August run for that reason, and the live page is what to cite. It has moved again since this was written, which is the point rather than an embarrassment: the live index carries the current count, and this post is left as the dated reading it always said it was.
• Six of the seven registry categories had scored brands on the 2 August 2026 run. Supplements had none, so nothing in this post describes it. That is a statement about the run this post reports, not a standing property of the index.
• No brand pays to be measured or to rank. There is no paid inclusion, no paid placement, and no way to buy a position.

The method is versioned and written down at the methodology page: what each collector does, what the eligibility floor is, and what gets retired when the rubric changes. If a number here is wrong, that page is where you would go to prove it, which is the point of publishing it.

The uncomfortable finding is not that some brands score badly. It is that scoring zero is invisible from the inside. An unmentioned brand produces no signal in anyone's analytics: no session, no bounce, no drop in a chart. Nothing arrives to tell you it did not happen. The only way to know is to ask the assistants and count.