Meta has handed its users a word that retail never built. Since February, a Threads post beginning “Dear Algo” has instructed the ranking system in plain English, and the instruction holds for three days. It can ask for more of something. It can equally ask for less. Shopping assistants have no answer to the second. The industry built recommendation without ever building refusal.

The interface was invented by the complainers. Threads users had been scolding the algorithm in public since the platform launched, without expecting a reply, and Meta turned the complaint into a feature. Its canonical form is a refusal: “dear algo, stop showing me posts about sick pets.” No fashion retailer has built the field that sentence would go in.

In June the request went private. Users now choose one, three, or seven days, and nobody else sees the ask; Threads passed 500 million monthly users in the same announcement. The refusal has stopped being a joke. It is a setting.

The trade reading treats this as a story about social platforms. TechCrunch filed it under platforms handing power back to users, and mentioned commerce nowhere. The more useful fact is grammatical. A sentence an ordinary shopper could say now moves a production ranker, which deletes “technically infeasible” from every assistant roadmap being signed off this quarter. Every product team that called it out of scope now has one fewer excuse.

Most shopping assistants listen for yes and treat anything else as a question not yet answered.

Retail is not without a no. It has the thumbs-down and the “not interested” tap, each of which kills a single product and then forgets. A standing instruction is a different object. It governs a category rather than an item, and the customer sets its expiry. Nobody has shipped the version that remembers for a week, because nobody has decided it’s worth building.

The benchmarks confirm the deafness rather than excuse it. Shopping Companion Bench tests whether an agent can hold a stated requirement across sessions, drawing on a pool of over 1.2 million real products; GPT-5 finishes below 70%. That is a real catalogue, not a toy benchmark. The authors trace the shortfall to two failure modes: agents hallucinating a preference they were never given, and agents that never check the product against what the shopper actually said.

The research version of the Threads idea exists and points the same way. CTRL-Rec steers a conventional recommender with plain-language requests, though its user study ran on nineteen Letterboxd members rating films rather than shoppers buying clothes. The transfer is not automatic. But a wardrobe holds more exclusions than a watchlist ever will, and each one is a sale somebody declined to make. Retail’s own data would make the case obvious, if anyone assembled it.

The mechanism is not the hard part; Meta assembled it out of a running joke. The hard part is commercial. An assistant that can be told “less” can be told to sell less, and no 2027 roadmap has been funded on that sentence.