As of November 9, 2026, Adobe's holiday tracking puts the share of products an AI agent can read cleanly at 66 percent, and finds 41 percent of items showing inconsistent prices across channels AI Best Practices: retailers must optimize product data for AI agents. Both numbers matter for price monitoring, but the second one is the harder problem. A missing field is a gap. A price that differs between the feed, the page and the agent answer is a contradiction, and contradictions are what get quoted back to shoppers.
Where the discount disappears
A discount is not one number. It is a base price, an eligibility rule, a time window and a channel scope, and the clause that carries all four is usually written for humans. An agent receives the base price and, if it is lucky, a promotional flag. The conditions that decide whether a shopper actually pays the discounted amount are frequently absent, and an agent that cannot confirm a condition will describe the safer, higher number instead.
That is why promo periods produce the widest spread. Outside a promotion, feed and page usually agree because nobody is editing them under time pressure. Inside a promotion, three teams update three surfaces on different schedules, and the agent reads whichever one it reaches first. The gap is not a pricing decision, it is a synchronisation failure that only becomes visible when someone asks an assistant for a price.
The 41 percent gap
Four in ten items carrying a different price depending on where they are read is a large number for a channel that is still treated as experimental. The practical consequence is that price monitoring now needs a fourth column. Comparing a marketplace listing against a brand site used to be enough. Add the agent answer, and the question changes from "which is cheaper" to "which one is the source of truth".
Meta's numbers explain the pressure. The company said more than 10 million people in the United States and Canada set up Muse within six weeks of launch, that shopping is the most common use, that more than a quarter of weekly users shop with it, and that the average purchase, travel included, is above $200 ACCESS Newswire: ReFiBuy launches Meta Muse support. A channel with that basket size is not a test surface, and a wrong price quoted there is a commercial event.
Clauses an agent cannot parse
Spend thresholds, bundle requirements, member-only eligibility and stacked coupons are the clauses that vanish first. They survive in an image, in a footnote, or in a terms page that no crawler prioritises. The result is a shopper told the pre-discount price by an assistant and the discounted price at checkout, which reads as a bait-and-switch even when the promotion is entirely legitimate.
Three surfaces, three answers
Auditing a single SKU across the three surfaces produces three different answers often enough that the pattern deserves to be logged as a routine rather than treated as an isolated exception. The table below is the working format used for that comparison, kept short enough to be filled in by hand during a store review and repeated weekly without extra tooling.
| Surface | What it shows | Common defect | Monitoring action |
|---|---|---|---|
| Product feed | Base price and promo flag | Flag set, condition missing | Add eligibility field, re-check daily |
| Product page | Price with footnote terms | Footnote not in structured data | Move terms into markup |
| Agent answer | One number, no conditions | Quotes base price during promo | Compare weekly, log divergence |
The middle row is where most teams stop, because the page looks correct to a person. The third row is where the shopper actually meets the brand. Between them sits a rendering step that quietly drops anything not in a field, which is why the fix belongs with merchandising rather than with whoever owns the assistant integration.
Labels are not the same as accuracy
Marketplaces have started asking that AI-generated titles and descriptions be submitted separately and labelled, which solves disclosure. It does not solve agreement. A labelled description that contradicts the feed price is still a contradiction, and disclosure does not repair it.
Auditing promo data
The audit can be scoped to the SKUs that carry the most promotional weight rather than the whole catalogue. ReFiBuy's guidance for the holiday peak is to bring the top 5 to 10 percent of SKUs to shopper-context enrichment first, which is a reasonable prioritisation rule for price work as well.
FAQ
Why does a price difference matter if the promotion is legitimate?
Because the shopper compares what the assistant said with what checkout charges. A correct promotion quoted incorrectly still reads as a broken promise.
Should price monitoring cover AI answers as a separate channel?
Yes, with its own column. The answer is what many shoppers see first, and it is generated from whichever surface the agent reaches, not from the price the brand intends.
Which fields close most of the gap?
Eligibility conditions, time windows and channel scope. Base price is usually present; the rules that decide who gets it usually are not.
How often should the comparison run?
Weekly in normal periods and daily during a promotion. Divergence that lasts longer than a day is usually a process gap rather than a data glitch.
