Profound's Own Report Proves the Wrong Point

“Christmas in July: How Retailers Win the AI Shelf” is careful, well-sourced work. Read closely, its own numbers argue against its own headline

Profound published a report this month, “Christmas in July: How Retailers Win the AI Shelf,” on how retailers win what it calls the AI shelf. It is careful, well-sourced work, built on a real, large dataset. It is also, read closely and in its own words, the clearest evidence yet for a claim the report does not make: that presence and recommendation are different things, and most of the industry, Profound included, is still only measuring the first one.

The Headline Is a Confession

Profound's own framing is that most AI visibility opportunities occur before the buying decision. Presented as a useful finding, worth reading literally: the entire dataset and playbook live on the pre-decision surface. That is the exact line we have spent this past year measuring across, presence versus decision, and Profound's report plants its flag entirely on one side of it without appearing to notice there is another side at all.

The word doing the most work in the report is recommend. What the data actually shows is which offers get cited most consistently at the first prompt, correlated with product page features. That is presence. Calling it recommendation is not dishonest, it is the natural vocabulary of a discipline built to measure mentions, but it is the single lexical slip that separates Profound's entire category of tooling from the question that actually determines revenue.

Profound's own funnel data shows the overwhelming majority of prompts in their dataset are not purchase moments. They measured the research phase and called it the decision.

Their Own Data Disproves the Single-Prompt Shopper

By Profound's own numbers, roughly two-thirds of prompts are research and comparison, and only around one in ten reflect a ready-to-buy moment. The dataset built to describe AI shopping is overwhelmingly composed of mid-funnel interrogation, not purchase decisions. This is corroborated independently: real ChatGPT sessions run to roughly twenty messages per weekly user, never a one-prompt shape, and the ICLR 2026 Outstanding Paper Award-winning study on multi-turn language model reliability found a 39% accuracy drop between single-turn and multi-turn evaluation, with models that commit to an early, incorrect assumption rarely recovering across subsequent turns. Retail decision-making is multi-turn. Profound's own funnel split says so, even where its framing implies otherwise.

A Method That Cannot See What It Is Measuring

Profound's dataset is built from 200,000 commercial prompts sampled from a single week, each classified in isolation. That design choice matters more than it first appears. Classifying prompts one at a time means the instrument cannot tell that two prompts came from the same shopper narrowing the same shortlist across a real conversation. Shopping looks single-prompt in this report not because shoppers behave that way, Profound's own funnel data says they don't, but because the method discards the connective tissue between prompts by construction. The arc of a real buying conversation is invisible to an instrument that never captures the conversation, only its fragments.

What This Confirms, and What It Does Not

None of this is a claim that Profound's data is unreliable. It is a large, real, carefully gathered dataset, and it independently confirms something worth having confirmed by a source other than ourselves: AI-driven shopping is real, growing fast, and already reshaping category discovery. That is useful, corroborating evidence for the underlying shift, sourced from a company with every incentive to describe that shift in the most favourable light for its own product.

What the report cannot do, by its own design, is answer the question its language implies it has answered. It can tell a retailer what people are asking. It cannot tell them how many turns a shortlist survives, or at which turn a brand gets quietly dropped, because it never captures a connected conversation in the first place. That is not a criticism of Profound's work. It is a precise description of where its own frame ends, stated in the report's own numbers, not inferred from outside them.

The Distinction Worth Holding

Found and recommended are not the same claim, and the difference is not pedantic. A brand can be the most-cited answer to a hundred thousand first prompts and still lose the overwhelming majority of the real, multi-turn conversations that actually end in a purchase decision. Our own research puts a number on that gap directly: 87% of brands present early in an AI buying conversation are displaced before the model's final recommendation. Profound's report does not contradict that finding. Read on its own terms, it is additional evidence for it, gathered by a different team, for a different purpose, and arriving at the same edge of the same unmeasured space.