The Shopper Is No Longer Human. Who Is Keeping Score?

More and more, the customer doesn’t choose the product directly. Their AI assistant does. Two studies published this month explain how that choice gets made, and why many brands are measuring the wrong part of it.

Oliver Wyman Forum surveyed close to 13,000 consumers across 15 markets on how they shop with AI. At the University of Toronto, Mahe Chen and Nick Koudas spent 90 days recording how ChatGPT, Gemini and Perplexity answer commercial questions, then replayed those questions to observe what happens inside the engine.

Taken together, the two studies describe three gates every brand now has to pass. The first is access: can the AI retrieve the brand at all? The second is choice: once retrieved, will the model pick it? The third is execution: can the purchase actually go through? Most of today’s optimisation work addresses only the first.

The demand side has already moved

Oliver Wyman found that a third of consumers now start shopping with AI, second only to search engines. When recommendations conflict, consumers say they would follow the AI over friends, family or the brand itself. The most telling figure is that 42% of AI shoppers begin by describing a need rather than naming a brand. For a growing share of purchases, the agent decides which brands enter consideration before the customer has named one.

The firm’s conclusion is that fit and evidence decide the AI shortlist. It calls the resulting discipline shopping agent optimization.

Inside the machine, memory comes first

The Toronto study complicates that conclusion, and its most important finding concerns what the model already knows. In effect, every model carries a memory of brands built during training, and that memory shapes recommendations before any fresh evidence is retrieved.

The effect shows up at every step. When a brand is part of the model’s prior knowledge, ChatGPT writes that brand into its own searches 67% of the time, against 11% otherwise. In 46% of cases where ChatGPT named a brand the user had not asked about, no cited page supported the mention.

In controlled tests with identical product facts, revealing the brand names reordered the rankings. Spectrum fell 33 points once named. When the researchers told the model that the names had been assigned at random, the same product profile still reached the top three 80% of the time under one name and 39% under another.

Retrieval comes second, and it follows its own logic. Every page ChatGPT cited came from its own search. That search explained 32% of the variation in citations, against 2.8% for every page feature combined. The engines rewrite each question into their own sub-queries, and 97% of the pages ChatGPT retrieved were absent from Google’s top 50. A strong position in classic search says very little about a brand’s position here.

The standard optimization playbook comes last. FAQ blocks, structured data, summary boxes and author boxes appear to help when popular websites are compared with obscure ones. When pages within the same website are compared, none of them has a measurable effect. These practices simply cluster on sites that get cited more for other reasons.

Discovery markets and reputation markets

The two studies can both be right because they describe different markets. The Toronto team replicated their work among small software vendors that the models barely know. There, the model’s memory counted for little. A third-party page naming a vendor raised the odds that ChatGPT named it by a factor of 39.7, against 4.2 for the large consumer brands.

These are two distinct economic regimes. In discovery markets, brands are judged on whatever evidence the engine’s search happens to find. In reputation markets, brands are judged largely on what the model already believes about them. Unknown brands compete for retrieval. Known brands compete for belief.

Oliver Wyman’s advice about fit and proof suits the challenger in a discovery market. For an incumbent in a reputation market, the larger risk sits in the model’s beliefs about the brand. No product feed or schema markup reaches those beliefs, and they can count against a brand even on identical facts. The Toronto authors hypothesise that changing a model’s internal representation of a brand could prove the most effective intervention. They stress that this has not yet been tested experimentally.

The three gates

Putting the two studies on one timeline shows where the sale is actually lost.

A brand fails at access when the engine’s search never retrieves it. This is the gate that generative engine optimisation addresses, and the Toronto evidence suggests most of its recommended tactics do little even there.

A brand fails at choice when it is retrieved but the model prefers a competitor, often because of what the model believes about the brand. For established brands, the Toronto evidence suggests this is where the largest effects sit. It is also the gate least measured.

A brand fails at execution when it wins the recommendation and the transaction breaks down. Oliver Wyman found that 79% of simulated agent journeys hit a barrier after the recommendation had been made. Shopping agent optimization addresses this gate.

What follows for brand owners

The cited set moves quickly. Ten weeks on, only 14% of ChatGPT’s citation pattern in the Toronto data was still shared, so a one-off audit describes a market that has already moved. And because the weight of the model’s memory differs by brand and category, no brand can know whether it is competing in a discovery market or a reputation market without measuring it.

AIVO’s own data points the same way. In “What a Citation Is Worth” (v1.1), an AIVO study led by co-founder Paul Sheals, identical prompts were repeated three times across 1,800 brand-level groups. The citations behind the answer changed 95.8% of the time. The brand the model finally chose changed in only 14.6% of groups.

The Toronto team saw the same churn from the outside: two ChatGPT runs of the same question on the same day shared only 15% of their cited pages. When the evidence turns over almost every time and the decision rarely does, the decision must rest on something steadier than the evidence.

For established brands, the Toronto findings suggest that something is the model’s memory of the brand. That is why the choice gate has to be measured on its own terms, and why counting citations tells a brand so little about whether it will be chosen.

Most optimization today measures whether a brand can be found. These studies suggest the larger commercial question is whether it gets chosen. AIVO was built to measure the choice: whether a brand reaches the recommendation, where in the decision it drops out, and whether changing the model’s understanding of the brand changes the outcome.

Sources: Oliver Wyman Forum, “Agentic Commerce Is Changing How Customers Are Won,” The AI Agenda, Chapter 3 (2026). Chen, M. and Koudas, N., “What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO),” arXiv:2610.15033 (October 2026). The Toronto paper is a preprint that has not been peer reviewed, and it used the smaller API models behind each engine and US-market questions. The authors of the Toronto paper are affiliated with ktau, a commercial AI visibility company. Their findings should be read in that light.

AIVOCORP