Retail has the cleanest feedback loop in software. A change either sells more or it does not, and you find out this week. That makes it a good place for model work and a bad place for hand-waving.
The catalogue is the bottleneck
Thin titles, missing attributes, three suppliers describing the same product three ways. Enrichment models fill the gaps and normalise the vocabulary, which improves search, filters, feeds and ad relevance all at once. It is unglamorous and it is usually the highest-return AI project a retailer has.
Search that understands intent
Keyword search fails on the queries with the most buying intent — the descriptive ones, the ones with a constraint in them. Semantic retrieval on top of a clean catalogue fixes a large share of those, and the ones it cannot answer are worth reading, because they are your merchandising gaps.
Support that can see the order
Most support volume in retail is three questions: where is it, can I change it, can I return it. An assistant that reads order state answers them in seconds. One that cannot is a worse FAQ page.
Where a model is the wrong tool
Recommendations from behaviour data, price rules, stock logic — these are usually better as ordinary systems. If a query answers it, a query is faster, cheaper and explainable. We will tell you that before we quote it.
Speed still decides revenue
None of the above is free at the page level. Every model call in a customer-facing path gets a latency budget and a cached fallback, because a slower page loses more than a smarter one gains. That argument sits with analytics and CRO, where the measurement lives.
Feeds and channels
Better attributes mean better feeds, and better feeds mean cheaper acquisition. The gain shows up in performance advertising and in technical SEO before it shows up on the product page.
Full picture: the AI map.