ChatGPT Got Better at Shopping. Will it Recommend Your Store?

OpenAI updated GPT-5.5 Instant this week, and the change deserves a second read from anyone who sells online. The default model inside free ChatGPT is now better at understanding the intent behind a question, with OpenAI pointing to improvements in shopping results, local recommendations, and what it labels complex constraints. Paying subscribers received it first, while free accounts gained access on June 25. The company published no benchmarks to substantiate the shopping claims, a gap VentureBeat flagged in its coverage.

A capable model that reads intent better raises the bar for your catalog rather than lowering it. As an assistant grows sharper at interpreting what a shopper genuinely wants, it depends more heavily on whatever your store can communicate about each product. The question arrives more specific. The evidence required to answer it has to become more specific too. Plenty of catalogs were engineered for an entirely different purpose, and they fall silent at the precise moment the assistant needs them to speak.

A smarter question meets a thin answer

Picture the kind of request OpenAI now parses more reliably. A shopper wants a jacket that survives a rainy commute, looks appropriate in the office, and compresses into a weekend bag. The model interprets that beautifully. Then it queries your store and receives four fields: a title, a color, a price, and an availability flag. Navy, one hundred twenty dollars, in stock. The assistant understood the shopper completely. Your catalog could not characterize the jacket well enough to earn its place in the recommendation.

Earlier this year I ran the same shopping query against six Shopify merchants through the agent protocol, and the plumbing performed cleanly every time. Every store returned tidy inventory data. Every store also withheld the reasoning a person relies on when choosing between two comparable jackets. It never says how the jacket fits, when it becomes the right pick, how it survives a rainy week, or what deserves a recommendation once the requested size disappears. None of that travels inside a standard feed. A more capable model amplifies the absence, because it finally knows enough to request the very details the feed abandoned.

The source-of-truth problem, moved to the shelf

The most valuable line in VentureBeat's coverage barely mentioned shopping. The spring release of GPT-5.5 Instant introduced a feature called memory sources, designed to show users which past chats and files shaped a personalized answer. Inside enterprises, those model-reported summaries repeatedly clashed with the deterministic logs of vector databases and retrieval pipelines. Teams ended up holding two competing accounts of what the system had genuinely consulted. Someone still has to decide which version counts as authoritative.

Retail confronts the identical question, only closer to the customer. When an assistant recommends your product, swaps in a substitute, or explains why a piece fits, it reasons from some underlying source. That source varies. It might be a scraped page, or a thin description written to rank in search. It might be a review snippet, a stale marketplace feed, or a rival's framing of your category. Or it might be something you authored deliberately, the entire premise behind context engineering.

Author the source, or let the model guess

That authored source is what we call brand truth at Sartorial. It carries the reasoning a feed omits. It names who a product suits and when, what to offer once the first choice sells out, the policies you stand behind, which item belongs in front of which request, and the voice you want carried into every conversation. A feed reports what sits in the catalog. Brand truth tells the assistant why to choose it, the reasoning an assistant weighs the moment it decides.

A fair objection: if the models keep getting smarter, won't they eventually infer all of this on their own? They will infer something. The trouble is that an assistant lacking fit guidance simply guesses the fit, one lacking substitution logic grabs the nearest lookalike, and one lacking policy data stays quiet about the guarantee that distinguishes you. Inference fills the vacuum regardless. What you author determines what fills it.

This is a modest update, another incremental step in the steady refinement of the default model that most people already use without much deliberation. That modesty is exactly why it matters. The assistants shoppers consult every day keep getting quietly, steadily better at understanding them, and every increment raises the standard for what your store has to say in return. Presence in the catalog keeps you eligible. Being understood by the assistant earns the recommendation, and authoring that understanding falls to you. If you want to see what an AI shopping assistant can reason about your store today, and where it runs out of road, I would be glad to talk.

---

Sumit Jagdale is the founder of Sartorial.