This article originally appeared on Forbes

‍

There is a moment in almost every considered online purchase when the browser tab closes without a sale. Not because the shopper decided against the product, but because one specific question—the question that mattered most—didn't get a straight answer.

For most of e-commerce's history, that moment happened on product pages, in review sections or on Reddit. Now, increasingly, it happens in a conversation with an AI shopping agent. More than half of consumers have already replaced traditional search with generative AI for product recommendations, according to Capgemini's consumer trends research.

These are not casual browsers. They are active buyers using AI to answer the questions that determine whether they purchase. What they are finding, consistently, is that the agent can tell them about a product, but it struggles to advise them on it.

Discovery Questions Versus Decision Questions

Two types of question define the AI shopping journey. Discovery questions are broad: "What are the best running shoes for flat feet?" The agent handles these well, surfacing options and narrowing a consideration set.

Decision questions are specific: "Does this serum actually work for combination skin, or is it better for dry types?" Or, "What do people who've owned this treadmill for a year say about the motor?" These are the questions a shopper asks when they are already in consideration and need one more piece of information to commit.

Ask an agent a decision question and you will typically get something like: "Many users report positive results, though experiences can vary. Some reviews suggest it may perform better for drier skin types."

Count the hedges. "Many users report." "Experiences can vary." "May perform better." Each phrase signals that the agent found the product, but cannot commit to an answer specific enough to close the sale.

The video player is currently playing an ad.

The consequences are well documented. Forrester research found that 55% of shoppers will abandon a purchase if they can't find quick answers to their questions. The AI hedge is that moment of unanswered uncertainty, arriving in a new place.

Why Agents Hedge

The hedging is rational, not accidental. The defining vulnerability of large language models is hallucination, or generating confident-sounding claims not grounded in verified evidence. For a consumer making a purchase decision, a hallucinated product claim—a moisturizer incorrectly described as suitable for sensitive skin or a supplement given false efficacy claims, for example—is a trust-destroying event for the agent, not just the brand.

Well-designed agents apply an explicit confidence threshold. When available evidence is thin, unverified or inconsistent, they default to hedged language rather than risk a confident wrong answer. The agent is doing exactly what it should when it lacks sufficient verified signal.

In other words, the problem is not the agent's behavior, but the quality of evidence available to it.

What Agents Want To Be

AI agents don't want to be search engines. They want to be consultants. Right now they can find your brand, but they can't advise on it.

The gap between finding and advising is evidence. The Capgemini study also found that 53% of consumers have already made a purchase based on a generative AI recommendation, but that number is held back by the frequency with which agents retreat into retrieval rather than recommendation.

The technology is capable of confident, specific advice. The evidence to unlock it, for most products and most brands, doesn't currently exist in the right form.

Why Current Sources Fall Short

Brand-generated content is abundant, but agents treat it as advocacy, not evidence—and correctly so. A brand asserting its product is "suitable for all skin types" is not independent testimony.

User-generated content from forums and social platforms presents verification problems agents cannot resolve. There's no way to confirm the commenter actually used the product, whether their profile matches the shopper's or whether the post is even genuine. Three-line Reddit comments provide nothing close to the specific, documented testimony an agent needs.

Editorial content appears authoritative, but agents are calibrated to its opacity. A best-of list may reflect genuine testing or paid placement, and shoppers sense this. SOCi found that 95% of consumers identify AI as their least trusted source for final purchase decisions. Consumers know that agent responses are only as good as their sources.

What Good Evidence Actually Looks Like

To make confident recommendations, AI needs:

• Verified Source Identity: Real, authenticated individuals whose relevant attributes—skin type, use history, purchase verification—can be confirmed.

• Documented Experience: Claims grounded in sustained, time-bound use rather than first impressions.

• Structural Specificity: Testimony broken into answerable claims—not "I loved this," but "After eight weeks of daily use with combination skin, I saw reduced T-zone oiliness without dryness on the cheeks."

• Transparency About Any Commercial Relationship: Creator content is not disqualified by a brand partnership, but the relationship must be disclosed, and the voice must be authentically the creator's or customer's own. An agent can work with disclosed creator testimony. It cannot work with claims that have been scripted or controlled by the brand, regardless of who delivers them.

When those conditions are met, the agent's calculus changes. It has something it can stand behind. The hedge resolves.

The Competitive Implication

Most brands tracking AI performance are measuring visibility, or how often they appear in AI responses. That number can look reassuring while the agent hedges on every decision question that actually drives purchase behavior. Only 2% of consumers currently act on an AI recommendation without additional verification, which shows the evidence gap. Visibility and getting recommended are not the same metric.

The brands that grasp that distinction early and make the evidence available will hold a structural advantage as AI becomes the dominant surface for considered purchase decisions. The decision question is where the sale is won or lost. Right now, for most brands, the agent is losing it.

‍