When AI Recommendations Get Prices and Stock Wrong
An assistant that quotes a stale price creates a disappointed checkout. How stores should ground recommendations in live price and stock data.
Shoppers increasingly meet a store through an assistant summary: a recommended product, a price, and a claim that it is in stock. When any of those facts are stale, the shopper experiences it as the stores mistake, even if the assistant invented the detail from an old crawl.
This guide covers where recommendation errors come from, what stores can publish to reduce them, and how to handle the shopper who arrives holding a price the store never offered today.
Stale facts have identifiable sources
Assistants reuse what they can read: cached pages, feeds, structured data, and third party listings. If a stores own feed updates daily while competitors update hourly, the stores price is the stale one in the comparison. Price and stock accuracy is partly a publishing hygiene problem, not only an AI problem.
Variant pricing is a common trap. An assistant may quote the smallest size or a sale variant as the products price, technically sourced and practically misleading. Clear variant pricing in structured data, with ranges where prices differ, reduces the confident wrong answer.
Publish price and stock like you mean it
Keep product feeds current, include availability per variant where it differs, and date promotional prices with start and end in machine readable form. A promotion without an end date in the data lives forever in summaries, long after the banner comes down.
Discontinued and sold out items need handling, not silence. A page that states sold out plainly, with a return date or an alternative, gives an assistant something accurate to say. A vanished page invites it to quote whatever copy it kept.
The arrival moment: honour the conversation, not the error
A shopper quoting an assistants price is not trying to trick the store, they believed a confident summary. Support scripts should acknowledge the mismatch, show the live price and where it is published, and have a defined goodwill path for close cases. Arguing that the AI was wrong, while true, wins nothing.
Track these arrivals as data. Repeated mismatches on one product usually point to one stale source, a feed, a listing, a marketplace copy, that can actually be fixed once it is identified.
Measure recommendation accuracy like a channel
Sample your own catalogue through the assistants shoppers use: ask for prices, stock, and delivery times on top products monthly, and compare with live data. Score the errors by type, stale price, wrong variant, phantom stock, and fix the largest class at its source.
Stores that treat assistants as a channel with a data quality budget get steadily more accurate representation. Stores that treat each error as an AI hallucination get the same errors monthly, because the source never changed.
Related reading
If this topic is useful, these earlier guides go deeper on connected questions: When AI Search Summaries Rewrite Your Product Claims, The Handoff Problem: When AI Chat Should Give Up Gracefully, and Product Feed Quality Is Now an AI Shopping Problem.
Frequently asked questions
Why do AI recommendations quote wrong prices?
They reuse readable sources: feeds, cached pages, structured data, and listings. Stale feeds, undated promotions, and unclear variant pricing in the stores own data are frequent causes. The assistant is confident, but the source was old or ambiguous.
What can a store publish to reduce errors?
Current product feeds with per variant price and availability, promotions with machine readable start and end dates, and honest sold out pages that state status and alternatives. Accuracy improves when the stores own data is the freshest source available.
How should support handle a shopper with a wrong quoted price?
Acknowledge the mismatch calmly, show the live published price, and follow a defined goodwill path for close cases. Then log the case. Repeated mismatches on one product usually trace to one stale listing that can be fixed.
How often should recommendation accuracy be checked?
Monthly on top products, by asking the assistants shoppers use for price, stock, and delivery facts and comparing against live data. Group errors by type and fix the largest class at its source feed or listing.