Study · July 2026

We asked AI what to buy in 21 DTC categories. Brand awareness had almost nothing to do with the answer.

Allbirds has been on the cover of Time. It went public. It has spent close to a decade, and a very large amount of money, making sure you know its name.

We asked ChatGPT and Gemini what to buy in its category. Allbirds scored 46 out of 100.

Ridge Wallet — a company you may well have never heard of — scored 81.

That gap is the finding. Across 21 direct-to-consumer brands we audited, how famous a brand is turned out to be a poor predictor of whether AI recommends it. And since a growing share of shoppers now ask an assistant before they buy, that gap is a revenue problem hiding inside a marketing win.

How we ran this

We make GetPick, a GEO agent for DTC brands. This study uses our own engine, and you should know exactly what it does before you trust the numbers.

For each brand we generated the questions a real buyer would type before purchasing in that category — not brand-name searches, but demand-side questions like “best machine washable sneakers for everyday wear”. We sent those questions to live AI assistants at audit time. No simulated prompts, no cached guesses, no modelled estimates. Then we recorded whether the assistant named the brand, or named someone else.

Limitations, stated plainly. The sample is 21 brands — enough to show a pattern, not enough to publish a law. Seventeen brands were audited across 12 buyer questions; four across 3, so their scores are lower-resolution and are marked below. Answers from AI assistants vary between runs and change over time; this is a snapshot, not a permanent ranking. And we are not a neutral party: we sell a product that fixes exactly the problem this study describes. Read accordingly — and note that we published our own score too, including the part that embarrassed us.

Finding 1: awareness and AI visibility are close to unrelated

The spread was wide: 31 to 88 out of 100, with a median of 65.

What it did not track was brand size. Necessaire — well funded, stocked in Sephora, the kind of brand that gets written about — scored 35. Brooklinen scored 88. Hedley & Bennett, whose aprons are on television, scored 31 and was named by AI on only 3 of the 12 buyer questions we tested.

Four of the 21 brands scored below 50. More than half landed between 50 and 69 — recommended sometimes, invisible often, which is arguably the most dangerous place to be, because nothing looks broken.

The reason is mechanical rather than mysterious. Traditional brand building buys recall in a human's memory. An AI assistant does not have your memory. It assembles an answer from what it can read. If your category expertise lives in a beautifully art-directed campaign and a paid-media budget, there may be very little for the model to work with.

Finding 2: someone else is being named, and it is often smaller than you

In 14 of the 21 audits, the assistant named a specific competitorin the brand's place. This is the part founders tend to find genuinely unpleasant, and it is the most useful signal in the data.

Hedley & Bennett lost to Tilit. Necessaire lost to Aesop. Versed lost to Cocokind. De Soi lost to Ghia. Spot & Tango lost to Ollie. Dagne Dover lost to Calpak. Tower 28 lost to Kosas.

Several of those winners are smaller companies with a fraction of the marketing budget. They are not winning on brand. They are winning because when a model assembles an answer about the category, their content is what is available to assemble.

That is bad news and good news in the same sentence. Bad, because ad spend does not defend this position. Good, because the thing that does win here is cheap to produce.

Finding 3: the losses are unusually fixable

Two things separated the top of the table from the bottom, and neither required a rebrand.

Machine readability. Some sites block or fail to serve the crawlers that feed AI answers. If the assistant cannot read you, no amount of content strategy matters — you have been eliminated before the question is asked. This is a robots.txt line and, increasingly, an llms.txt file. It is a fifteen-minute fix that is either done or not done.

Answering the actual question.The brands that scored well tended to have plain, crawlable pages that answer category questions in the buyer's words — comparisons, “best X for Y” framing, specifics about materials, use cases and trade-offs. The brands that scored poorly often had gorgeous sites that described the brand rather than the decision.

The table

* audited across 3 buyer questions rather than 12.

We audited ourselves, and it went badly in an interesting way

We scored 61, and AI recommends us on 9 of 12 buyer questions — respectable.

The problem was what it said about us. Asked to describe GetPick — formerly Citeable — the assistant called us “an intuitive no-code solution for local businesses.” The buyer questions it generated for us were about local restaurants and food brands.

The audit also caught a naming problem, which explains the signature change on this study: it is now signed GetPick, formerly Citeable. Our old name was already carried by two other players in this exact category — the same attribution failure we measure for everyone else — so we renamed. The data in the table is unchanged.

We do not sell to local businesses. We sell to DTC and e-commerce brands. The AI had us filed in the wrong category entirely — confidently, and in a way no dashboard metric would have flagged, because our visibility number looked fine.

This turns out to be the more important lesson in the whole study. Being cited is not the goal. Being cited in the category where your buyers are asking is the goal. A brand can score well and still be invisible to the people who would actually buy from it.

We have since published an llms.txt, declared our category and audience in structured data, and explicitly allowed the AI crawlers. We will re-audit and report whether it moved. If it does not, we will say so.

What to do with this

Find out whether AI can read you at all. Check that your robots.txt does not block GPTBot, ClaudeBot, PerplexityBot or Google-Extended. This is binary and takes minutes.

Find out who is being named instead of you.Not whether you are “visible” in the abstract — the specific competitor the assistant recommends when someone asks what to buy in your category. That name tells you what to write.

Publish the page that answers the question.Not a brand story. The comparison, the trade-offs, the “best X for Y” in the words your buyer actually uses.

Method, dates and per-brand results available on request. If you are one of the brands in this table and want your full audit, ask and we will send it.