Market Commentary

AI Product Recommendations: what happens to a brand's shelf space when a machine gives only one answer

Shoppers increasingly ask an assistant what to buy instead of scrolling a results page. This note reads across the first hard evidence on who gets named, who pays for visibility, and who referees the claim. It is written for anyone whose brand depends on being found.

By Commerce desk × Folka

1. One answer, not ten rivals

In September, Marketplace reported that a marketer built a website for a deodorant brand that did not exist, Morrowen, built around a single ingredient claim, and within about three weeks ChatGPT's browsing mode was ranking it above established names. The case is narrow — it worked in a thin category, on one assistant, for one kind of question — but it is a useful opening because it makes visible something that used to be invisible: a results page holds many options side by side, and a shopper can compare them. A chatbot's answer holds one name, maybe two, and there is nothing on the page to weigh it against. Call this the answer shelf: the single, narrow space a generative assistant hands a shopper in place of a results page, and the only shelf space a growing share of product discovery now runs through.

Bain's research, as Marketplace cites it, found close to three in ten shoppers already use AI for product recommendations or comparisons. The Morrowen case is a single data point; the scale behind it is not.

2. Gaming the answer is an arms race with no winner

The firmer evidence is a preprint by Chu and Hou (arXiv 2606.17443, revised 10 October), which tested identical skincare specs across GPT-4o-mini, Claude Sonnet and Gemini 3 Flash. On matched products, the better-known brand won every time. A rival needed as little as a tenth of a star's extra rating, or confident authority language — including fabricated clinical claims — to overturn that. The part that matters most is the competition test: once every brand in the set used those tactics, each one's measured gain collapsed from +0.802 to +0.007, and the one brand that refused to play gained nothing at all.

That dynamic punishes restraint and rewards whoever moves first, which is exactly the condition that turns a tactic into an industry norm. A lab result, though, is not the same as a live market: a model in a test is handed the tenth-of-a-star difference. A live assistant has to go and find it.

3. The lab rates; the wild talks

A second preprint, by Malthouse and colleagues (arXiv 2609.16304, 14 September), tested six models across five categories and found a different lever at work outside the lab. Category-only questions often left established brands out entirely. Prominence tracked search interest and online conversation more closely than it tracked conventional popularity measures. And simply adding a user's stated needs changed which brands got named at all.

If that holds, the lab's tidy ratings contest is the easy version of the problem. The harder version is sticky: a brand that is barely discussed has little to adjust, and the places that shape the conversation sit outside a brand's own control.

4. Small shops get a real share, but not a default

The one dataset on independent retail comes from a Vaer AI study commissioned for Lightspeed, a vendor that sells to independent shops — worth noting as a commissioned source, not a disinterested one. When ChatGPT's own web search named a store, it was a chain 97% of the time; small shops appeared in only 1 to 4% of its searches. Google's AI Overviews named no small or local store in 67% of its answers. Yet with search switched on, small shops still took around a quarter of the assistant's top picks, against 46 to 58% for chains.

That is not erasure, and it is not parity either. It is a default that favours scale at the point of discovery, with a narrower but real opening at the point of the final answer.

5. Three different relationships between the shelf and its owner

In the West, the paid layer still sits beside the answer rather than inside it. ChatGPT runs labelled, separate ads; advertisers can target hashed-contact lists of at least 25,000 matched users (Yahoo Tech, 15 October). OpenAI briefly charged merchants a checkout fee inside ChatGPT, then moved purchases back to merchants' own sites. Elsewhere the model and the shelf are the same business. On 11 May, Alibaba connected its Qwen assistant to the entire Taobao and Tmall catalogue — more than four billion products — with comparison, ordering and an automatic price tracker. In India, Google is testing purchases from Flipkart inside Gemini.

These are not stages of the same path so much as three live models sitting side by side. Which one a market ends up with changes who has to be paid, and who gets to referee, before a brand is ever named.

6. The claim has a law; the answer does not

A brand that fabricates a clinical claim on its own site already breaks the law in most large markets: the US Federal Trade Commission's rule on fake reviews and testimonials has carried civil penalties since October 2024. What has no equivalent rule yet is the assistant that repeats the claim as its answer. Europe has ruled on self-preferencing before — the European Commission fined Google €2.42 billion in 2017 for favouring its own shopping results, and the EU's Court of Justice upheld that fine in September 2024 — but that case concerned a company ranking its own products, not a model amplifying someone else's unverified claim.

Nothing in the record says which way this settles. What is already visible is that the gap exists, and that an answer shelf with no referee is a different risk environment to a results page sitting under decades of advertising and review law.

7. A confident answer is persuasive because it is alone

Part of why authority language works on a model is that it also works on the person reading the model's answer. A results page shows ten competitors, so a shopper compares; an assistant's answer shows one, so there is nothing to compare it against. Logg, Minson and Moore's 2019 research on algorithm appreciation found that people often weight advice more heavily when told it comes from an algorithm than when told it comes from a person. The same confident, credential-sounding phrasing that moves a model's output is landing, unfiltered, on the shopper reading that output.

That compounds the earlier findings rather than sitting apart from them. The same mechanics that let authority language move a model's answer also make that answer land harder on the person who receives it.

8. Our view

We think the Morrowen case is the least important piece of evidence here and the most repeated. The real story is structural: Chu and Hou show that gaming the answer shelf is a trap that resets itself, Malthouse and colleagues show that in the field, visibility tracks conversation rather than rating, and the Lightspeed-commissioned figures show that scale, not merit, sets the discovery default even where a real opening survives at the point of the final answer. Put together, this is not a ranking problem that better copy solves. It is a market where the reward for gaming resets to zero the moment a category escalates together, where the lever that actually moves prominence sits in forums and publishers a brand does not own, and where the only clean legal hook — fabricated claims — attaches to the speaker, not the assistant that repeats it.

The deeper misreading is treating the answer shelf as a new SEO problem, something to be optimised. It behaves instead like a market with no settled property rights: nobody yet owns the rules for who gets named, who pays for that, and who is liable when the name is wrong. The Qwen and Flipkart moves point to one resolution — the shelf absorbed into the retailer's own commerce, where the question of who controls visibility dissolves because the model and the store are the same business. The FTC letters on Amazon and Walmart's assistants, and the EU's Google Shopping precedent, point to another — liability extended to the amplifier, which would make the assistant's answer a regulated surface in its own right. Both are live. Neither has happened yet at scale.

We would weight the field evidence over the lab evidence and the structural incentive over the individual tactic. A brand chasing a tenth-of-a-star edge is solving yesterday's problem; the harder and more durable question is who is talking about a brand where a model reads, and who, eventually, is legally responsible for what the model says next.

9. Three shelves, not one

Three structures are visible in the evidence, and none has won yet. In one, the answer shelf stays an open contest, with ads beside the answer and visibility earned through conversation and reputation rather than ownership — the model ChatGPT's ad business and the Malthouse findings both point toward. In a second, the claim gets policed through liability extended from the brand to the assistant that repeats it, following the logic of the EU's Google Shopping ruling rather than the FTC's current claim-only enforcement. In a third, the shelf is absorbed into the owner's own commerce, as Alibaba has already done with Qwen and Taobao, and as Google is testing with Flipkart — at which point the question of who controls visibility stops being a market question and becomes an internal one.

Which of these a given market lands in depends on decisions still in motion — a Senate letter to the FTC, an EU court precedent being tested against a new kind of ranking, and a retailer's own choice about whether to connect its assistant to its catalogue. Worth watching without picking a side: whether liability law moves from the claim to the amplifier, whether more assistants follow Qwen's merged model outside China, and whether independent retailers' quarter-share of top picks holds as more categories get this kind of scrutiny.

How we read this

This note draws on a commissioned retail study (Vaer AI for Lightspeed, with Lightspeed as an interested vendor), two preprints not yet peer-reviewed (Chu & Hou; Malthouse et al.), contemporaneous reporting (Marketplace, Yahoo Tech), and one peer-reviewed behavioural study (Logg, Minson & Moore, 2019). Figures from the preprints and the commissioned study are measured within those specific tests, not market-wide; the Morrowen case is a single reported instance. The read on where the three forks lead is editorial judgement, not a projection. The one figure that would most change this read is a replication of the Chu and Hou escalation result outside a controlled test, on a live assistant's web search.

Grounded in Folka’s corpus: Cited, Not Ranked · The Model Native Brand · Ask, Don't Type · Display Neutrality