Marketplace Ratings: The Agent's Decision Input

Consider a composite scene. A US smart-lock brand rebuilds its DTC site around answer units; a chat assistant soon quotes it for a renter’s question. That week, a second assistant answers “renter, no drilling, under $150, which lock?” with three listings. The first carries “4.3 stars, 2,100 ratings,” then “buyers report the instructions are unclear and older frames may not fit.” The site never appears, and that complaint comes from a two-year-old review. The budget went to the site, not the backend.

A marketplace rating is not storefront decoration; it is behavioral evidence that travels with the SKU, and an agent uses it to gate candidates, justify a recommendation, and price the worst case; only the seller backend changes it.

Ratings travel with the item, not the store

The platform compresses item, seller, and fulfillment into one retrievable object: a question matches one listing, and one sentence carries stars, count, a complaint, delivery promise, and return policy. Your site is a separate description the model must reconcile at higher cost.

The two kinds of evidence are not equally cheap to fake. “Ships in 48 hours” is a claim; a late-shipment rate is sediment, because buyers voted with payments and returns. A returned parcel costs money twice, so the record outweighs the sentence.

The counter-example is a pattern: a trust page promising two-day dispatch, a late rate above target for three months, reviews saying “it took ten days.” The agent does not average that; it retreats to the conservative reading or another seller. Agentic commerce readiness maps the chain.

How an agent uses the numbers

Use	What it looks like	What you observe
Hard gate	Below a line on stars or count, no entry	Absent for months
Citation reason	Stars and count as “why this one”	“4.3 stars, 2,100 ratings”
Risk flag	A complaint becomes a stated downside	“Instructions unclear”
Ranking nudge	Ties broken by fulfillment	Behind peers

Teams push 4.3 to 4.5 and the answer still omits them: the gate and the risk flag block before the decimal matters. Thresholds are unpublished, so read the rows as an order of diagnosis, not a target.

Rating count is a sample-size problem: a 4.8 with 20 ratings loses to a 4.3 with 800 on “renter who cannot drill,” because a model needs enough specific detail to commit. Negative reviews surface for the same reason: they are more specific than your page.

Five metric groups, three speeds

Metric	What the agent judges	Speed	Common failure
Stars plus count	Baseline and sample size	Monthly	Five five-star ratings as proof
Complaint text	The worst case	Quarterly	“Sorry for the inconvenience”
Late, cancellation rate	Whether the promise holds	Ship cycle	No restock after a promo
Return rate and reasons	Wrong buy or bad product	Monthly	All filed as “other”
Dispute rate, refund time	Who backs the buyer	Monthly	Waiting for the platform

Fulfillment moves on a shipping cycle, stars on a quarter, return reasons slowest.

Example: one smart lock, three backend changes

Example — all figures invented.

Before: the listing promises “fast shipping, professional installation guidance”; handling time is two days; 200-plus “won’t fit” complaints sit under “other,” with no compatibility note on the listing. The agent cites it, then adds: “some buyers report installation difficulty; confirm your door frame size.”

Three changes. Fulfillment: one-day handling, “order by 2 pm ships same day” in the listing, an honest arrival window in the feed. Returns: “won’t fit” and “size mismatch” split out of “other,” plus a frame-thickness table. Reviews: the five-star blast becomes a day-20 question — “which door, and how long have you used it?” — and the two loudest complaints, an adapter plate for old frames and proximity pairing, go into the page.

Two months later: “4.4 stars; most reviews mention easy installation; older frames need an adapter.” The flaw remains, but it moved from unknown risk to known condition — which decides whether an agent shortlists you.

Fix what can be verified first

Order	Action	Owner	Passes when
1	Pull 90 days of backend metrics	Operations	Every item is a number
2	Fix the worst line, usually fulfillment	Ops, logistics	Late rate falls two cycles
3	Split “other” into reasons; change listing and page	Content	Top three reasons have fixes
4	Ask for the setting, not the score	Support	Reviews name a use case
5	Align site, listing, feed on timing and returns	Technical	Three SKUs, no contradiction

Step 2 comes first because it can be verified: stars take a quarter, fulfillment shows in a week, and agents weigh arrival above “4.3 or 4.4.” Invert the order and you get a prettier average and an unchanged answer. The listing half is product copy for agents; channel choice, marketplace versus DTC.

Where this breaks

The fields are closed. When a platform hides backend metrics, an agent reads only the public listing. Check which fields appear in answers, then fund only those.

There are too few reviews. Under roughly 50, the problem is “does anyone review this,” not “is the score high enough.” A thin sample earns no citation.

Baselines are category-specific. Electronics and home goods do not share a scale, and marketplaces change display rules. Treat specifics here as a September 2026 snapshot, not a spec.

You are tempted to manage the number. Incentivized reviews, cashback for stars, and bulk complaint deletion breach marketplace rules and consumer-protection law in the US and the EU. The cost is the listing, not a tenth of a star.

You sell mostly DTC. Best-of lists and third-party reviews quote marketplace ratings directly, so the layer is checked against your brand entity. Sample answers instead of one screenshot, as an AI visibility baseline does.

This month, put the 90-day numbers on one page, circle the worst line, and fix it across two shipping cycles. Sixty days later, ask the same question and read the sentence the agent uses: did the downside move from unknown to known, or are you still absent? This line continues on the authority and off-site hub.