AI search is becoming a decision surface

McKinsey’s October consumer research is a useful signal, not a universal forecast. Its June survey of roughly 21,000 consumers across 11 markets found that 53% had used AI for shopping in the prior three months. Among AI shoppers who described a recent purchase, AI chat or search was the top-ranked influence for 29%. The survey also found that standalone AI chat and search tools were used more often than retailer assistants during much of the journey, while retailer tools regained ground near checkout.

The relevant lesson is not that conventional search has disappeared. It has not. A 2026 Journal of Advertising Research study found heavier reliance on traditional search engines than on AI-powered ones. But it also found that people used AI-powered search more in advice-seeking situations and wrote more elaborative queries there. That is exactly where a company’s category, trade-offs, credibility, and fit tend to be judged—not merely located.

A separate large-scale field experiment on Meituan found that generative search increased purchases, reduced exploratory browsing and clicking, and concentrated evaluation within relevant categories and merchants. The setting is one platform and cannot establish the same outcome for every category or market. Still, it gives a concrete mechanism for why the interface shift matters: synthesized answers can change how a person formulates the problem and which attributes enter the evaluation before the person ever opens a brand page.

Most visibility programs were built to answer a simpler question: did the market encounter us? AI answers demand a second question: what did the system help the market conclude? A company can be named in an answer and still be framed as expensive, narrow, risky, interchangeable, or appropriate only for a use case it does not want to own. Conversely, a company can be absent from a generic category query but appear consistently when the question includes the customer problem it is actually built to solve.

That is why a useful AI-search review should preserve the question, not just the result. Record the buyer or practitioner question in full; the response’s relevant claim; the companies, sources, and attributes it chose; the caveats it raised; and the date, provider, and interface used. Then compare that record over time. The objective is not to force a flattering answer. It is to identify whether the market-facing explanation is accurate, supported, and aligned with the decision the customer is trying to make.

This is a communications discipline as much as a search discipline. The raw material is distributed: public product information, independent reporting, customer-facing documentation, expert commentary, reviews, and the language people use when they describe a problem. A strong answer may draw on any combination. A company that treats every unfavorable answer as a technical ranking defect will miss the harder possibility—that the available evidence supports the answer it received.

Start with a bounded question set rather than a large pile of prompts. Use questions that map to real decision moments: what problem is the buyer trying to solve; what alternatives are being compared; which criteria are likely to decide the choice; and what evidence would make the recommendation credible? Keep those questions stable long enough to observe movement, while labeling any change in provider, product surface, geography, or audience assumptions.

For each question, assess five things. Fit: is the company present when the question genuinely matches its role? Frame: what category, strengths, and limitations does the answer assign? Evidence: what sources or assertions support the explanation? Alternatives: which competitors or substitutes are placed beside it, and on what criteria? Risk: is there an inaccurate, stale, or unsupported claim that could influence a consequential choice? This converts an unstable answer into a reviewable observation rather than a vanity metric.

The counterargument is important. These interfaces are dynamic, model outputs vary, and a small prompt set is not a census of market perception. They should not be used to claim deterministic attribution or to replace research with customers, analysts, journalists, or communities. The better use is narrower: treat recurring answer behavior as a signal to investigate. If a pattern persists across defined questions and surfaces, examine the public evidence behind it and decide whether the right response is correction, stronger proof, clearer positioning, or no action at all.

AI search makes a familiar organizational weakness easier to see. Many companies can publish, monitor, and report, yet cannot reconstruct what they believed about an audience, what they changed, what the market then reflected back, and what they learned. Without that memory, each visibility review starts from zero and each answer becomes a new emergency.

The durable asset is a maintained record of the relationship between company claims, external proof, audience questions, and observed interpretation. That record helps a communications team distinguish a transient output from a material pattern. It also gives leaders a better decision: not “how do we appear everywhere?” but “where does the market’s explanation of us diverge from the evidence we want a buyer to use—and what, if anything, should change?”

AI-mediated search will not make every purchasing journey agentic, and traditional search will continue to matter. But the growing use of answer interfaces at the point of comparison means visibility can no longer be measured only as traffic or placement. When a system begins helping people formulate and evaluate the decision, the company needs to understand the explanation—not just count the exposure.