80 percent would let an agent pay, 59 percent still want to approve each order themselves, and that gap is where most marketplace trust strategies fail. As of November 15, 2026, the Worldpay 2026 Agentic Commerce Report is the clearest public read on how far consumers will delegate, while Google's Universal Commerce Protocol and OpenAI's merchant feed work define what a seller must publish to be readable at all. The practical question for a marketplace seller is narrower than the debate: which product fields decide whether an agent recommends an item, and which ones decide whether a shopper approves the order.
Trust splits into two decisions
The two numbers describe different moments. Letting an agent search, compare and assemble a basket is a research permission; letting it spend money is a financial one. Sellers that treat them as a single trust threshold build the wrong features, usually a faster checkout when the shopper actually wants a clearer confirmation screen.
Approval behaviour also explains why agent orders and human orders cannot share a fraud model. An agent that buys the right item in an unusual pattern looks like fraud to rules tuned on human sessions, and a wrongly declined order costs the same as a lost sale while also damaging the account's standing with the platform. Separating agent traffic in analytics is the cheapest step, and most sellers have not taken it.
What an agent reads before it recommends
Fields, not pages
An agent compares specifications, not banners. Size, material, compatibility and pack count have to sit in named fields with units, and the price and stock in the feed have to match the live page. When the feed says 129 dollars and the page says 99, the recommendation survives but the approval fails, which is the most expensive failure mode because the traffic was already paid for.
Review analysis as evidence, not decoration
Review volume and recency work as evidence density. An item with a thousand reviews, none from the past six months, reads as inactive; an item with forty recent reviews and a clear correction history reads as current. Sentiment scores summarise the same evidence, which is why a five-star average with recurring complaints about one specification still loses to a lower-rated item whose reviews are specific and recent.
| Signal | What it answers | How to record it |
|---|---|---|
| Spec completeness | Can the item be compared at all | Share of attributes filled with units |
| Review recency | Is the evidence current | Reviews in the last 90 days |
| Stock accuracy | Will the promise hold at checkout | Feed versus page mismatch rate |
| Approval rate | Does the shopper accept the order | Approved orders over agent sessions |
The fourth row is the one sellers cannot see today. Agent sessions, conversions, declines and cancellations need their own reporting line, because mixed into human traffic they disappear into averages that move for unrelated reasons. Consumer feedback arriving through an agent also needs a channel of its own, since a complaint typed into a chat rarely becomes a product review that a listing owner would ever read.
Review evidence still closes the sale
Corrections beat averages
A review that corrects a specification is worth more than ten that praise the delivery. Corrections point at a field to fix, and fixing it removes the objection from every future answer the agent gives. Tracking how often a corrected field appears in later reviews is a direct measure of whether the fix worked.
The confirmation screen is a trust feature
With 59 percent wanting per-purchase approval, the confirmation screen carries more weight than the checkout flow. Price, quantity, delivery window, return terms and the substituted item, if any, all belong on that screen in plain text. It is also the last place a seller can prevent a return, which costs more than the cancelled order it replaced.
Sentiment analysis on its own will not catch this. What catches it is a review analysis that separates specification complaints from delivery complaints, tracks how recent each review is, and routes every correction back to the field it belongs to. A marketplace that runs that loop monthly turns consumer feedback into feed hygiene, and feed hygiene is what an agent reads before it recommends anything.
None of this requires a new platform. Most of the work is catalogue hygiene on data the seller already owns, plus a reporting line that separates agent traffic. Adobe's holiday forecast, which expects AI referral traffic to US retail sites to rise 130 percent this season, suggests the volume will arrive before the reporting does, and the sellers who instrument it first will be the ones able to argue about attribution with evidence.
FAQ
Does the 59 percent approval figure mean agents will not be used?
No. It means shoppers delegate research faster than they delegate payment. Sellers should expect agent traffic to convert at a lower rate until confirmation is clear.
Which field causes the most failed agent orders?
Pack size and variant, because they are the fields most often implied rather than stated. A page that says large without a number cannot be compared by an agent.
Should agent orders use the existing fraud rules?
No. Patterns that look anomalous for humans are normal for agents, and false declines cost both the sale and the account standing.
What is the first metric to add?
Agent order decline rate. It is small, cheap to instrument, and reveals whether the risk stack is rejecting good orders.
How should review analysis change when agents shop?
Separate specification complaints from delivery complaints, weight recent reviews, and treat every correction as a product data fix rather than a customer service case.
