Fanatics Shop the Look Gets OpenAI Recommendations: What Changed

Fanatics has introduced an OpenAI-powered version of Shop the Look, recommending coordinated merchandise while shoppers browse a product. OpenAI's October 26 announcement says it is live on Fanatics.com and official NFL, NBA, MLB and NHL stores. The change puts AI inside a familiar shopping step rather than asking customers to learn another interface.

For someone already considering a jersey, the next question might be which cap or other item goes with it. That is a narrower problem than asking an assistant to choose everything they should buy. It also gives a product team a clearer place to evaluate whether its recommendations help.

How Fanatics Shop the Look works

According to OpenAI's account of the rollout, the system combines enriched catalog data, team information and inventory with Fanatics' merchandising rules. Suggestions use available products that fit the team and intended style. Looks refresh daily.

Consider the difference between finding similar items and building a useful combination. A second jersey might look very similar to the first one, but it does not necessarily help someone finish an outfit. A complementary recommendation has to understand the relationship between products, not simply rank them by resemblance.

Sports merchandise adds another constraint. Matching colors alone would be a weak basis for a recommendation if the resulting combination crossed team loyalties. A practical system needs a way to distinguish preferences it can explore from conditions it should preserve.

That suggests a useful engineering pattern for other retailers: let the model help interpret messy product information, then keep commercial constraints explicit. Whether an item is purchasable, compatible or appropriate for the category should be something the team can inspect and test, rather than an assumption hidden inside a fluent explanation.

What Codex did in the build

OpenAI says the teams used Codex to develop tools for catalog enrichment, compatibility checks, evaluation and look generation. The first 50 products were tested in under a day, followed by human feedback and an MVP approved that week. The runtime model is not named.

That distinction matters when evaluating the result. A coding agent helping engineers build a system and a model making recommendations inside that system are separate parts of the architecture. Crediting the development tool does not tell us the serving model, its latency or the cost of each shopping session.

The more useful lesson is the shape of the early test. A small set of outputs can expose obvious mistakes before a team expands coverage. For a jersey recommendation, a reviewer could check team alignment, product compatibility and whether the proposed combination makes sense as an outfit. Those checks create concrete feedback for the builders.

What the 10% engagement result means

After testing roughly 30,000 MLB products, the company reports 10% higher engagement than the previous experience. That is an engagement result, not a disclosed increase in sales. The announcement does not publish conversion, revenue, inference cost or detailed experiment methodology.

Those distinctions change how a business should read the case study. More interaction with recommendations could indicate that shoppers found them relevant. It cannot, on its own, establish whether more orders were completed, whether baskets became more profitable or whether returns changed.

Why the existing product page matters

For teams considering a similar feature, the first design question should be where customers already have enough intent to benefit from assistance. On a product page, the selected item supplies context. That can make the recommendation problem more focused than starting from an empty chat box.

The next evidence worth watching is whether this kind of assistance improves completed shopping outcomes after operating costs are included. A more appealing recommendation is a useful starting point; the durable value comes from helping someone make a decision they remain happy with.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams explain how their AI products work and turn practical product decisions into stories developers and founders want to read.

Book a 30-minute call.