OpenAI published its account of the Fanatics deployment on October 26, 2026, and the feature it describes is already running: Shop the Look is live across many sites in the Fanatics network, including Fanatics.com and the official NFL, NBA, MLB and NHL online stores. The page reports that the OpenAI-powered version of the module gets 10% higher engagement than the version it replaced.
What was actually released
Shop the Look is a merchandising module, not a chat window. A shopper who opens a jersey sees coordinated outfits offered on the product page, assembled from items that are in stock and matched to the team and to the style of the item being viewed. OpenAI models draw on enriched product data, team information and inventory; Fanatics merchandising rules keep the combinations compatible and relevant, and the sets are refreshed daily as catalogs change.
The build is the other half of the release. OpenAI and the Fanatics engineering team used Codex to write custom tools that enrich product and catalog data, find compatible merchandise, evaluate candidate recommendations and generate looks for each product. On the corporate side, OpenAI says Fanatics uses ChatGPT Work and Codex across Fanatics Corporate, Commerce, Collectibles and Betting and Gaming.
What the release actually consists of
- The shopper does not type anything: the suggestions appear on the product page, built from in-stock items.
- Codex was used to build the tooling around the catalog, not to answer shoppers.
- The first working version was tested on 50 products in under 24 hours, before any wider rollout.
- Published scope: Fanatics.com, the NFL, NBA, MLB and NHL stores, with roughly 30,000 MLB products in the recommendation test.
The timeline is worth reading closely, because it is where most of the claim sits. Testing on 50 products took less than a day. Fanatics reviewed the looks and gave feedback straight to the builders; after a couple of days of iteration the company approved the minimum viable product that same week, and the approach then moved into production and onto thousands of products.
What the claim rests on
The page publishes four checkpoints, and each one measures something narrower than a reader would assume.
| What OpenAI reports | The figure | What is not published |
|---|---|---|
| Engagement | 10% higher than the previous Shop the Look | Definition of engagement, test window, traffic, control |
| Testing scale | ~30,000 MLB products in recommendation testing | Whether those products were live, and the other leagues |
| Testing speed | under 24 hours for the first 50 products | Engineering hours, review load, cost |
| Approval | MVP approved the same week | Headcount, what was reused from existing code |
The 10% is a lift against Fanatics’ own earlier version of the same feature, not against a store that shows no recommendations at all. That is the right baseline for a retailer swapping one module for another, and it is the wrong number to quote as the return on an AI project. Nothing on the page says which metric moved, whether clicks on a suggested item, add-to-cart rate or revenue per session, nor over what period, nor on how much traffic.
Without a stated window and a control, the figure cannot be separated from seasonality. A launch timed to the fall sports calendar lands in the weeks when football and hockey merchandise sells hardest, and the page dates the release that way itself.
The 30,000 figure is described as products “included in recommendation testing”, which does not say that the whole catalog is covered, or that the tested looks ran live. The models powering the recommendations are not named, and no version is given, so the result cannot be tied to a model generation or checked against a later one.
After a couple days of iteration, the company approved the minimum viable product that same week.
What is not published
No model name or version, no definition of engagement, no test window, no control group, no traffic volume, no rate of irrelevant or wrong-team recommendations, and no cost per thousand recommendations. Only the MLB test is sized: the NFL, NBA and NHL stores are named as live, with no figure attached.
On this page
What it changes for teams building storefront recommendations
For a retailer, the transferable part is the order of work, not the model. Enrich the catalog data first, encode merchandising rules as constraints, generate candidates, evaluate them against those rules, and only then let the module face shoppers. Most recommendation pilots break at the data step and blame the model.
The second reusable piece is the size of the first test. Fifty products is small enough to review by hand and wide enough to expose a missing attribute or an empty size. It also puts the people who will judge the output in production in front of that output early, which is the likeliest reason approval took a week rather than a quarter.
The third piece is for the buyer, not the builder. A page like this is a template for the questions worth putting to any vendor who quotes it: which metric, over which period, against which counterfactual, on how much traffic, and at what cost per recommendation.
Our earlier coverage of vendor build claims points the same way. OpenAI’s forward deployed team described a two-month first build that turned into a one-day repeat rollout, and Block reported its patch time cut to 18 minutes with OpenAI’s security tooling. In each case the headline is a build-speed or before-and-after figure measured by the customer, not by an outside party.
The Neural Minute take
The number a buyer will remember is 10%, and it is the weakest thing on the page. What deserves attention is the pipeline: catalog enrichment, rules as guardrails, generated candidates, human review, then scale. That is a merchandising system with a model inside it, and it is the only part of this story a team can copy on Monday.
Our reservation is structural. Recommendation quality depends on the constraint layer, and here that layer belongs to Fanatics: the page says Fanatics merchandising rules keep the selections compatible and relevant. So the thing that protects the brand, a look that never mixes rival teams or sells an out-of-stock size, is exactly the thing the vendor’s case study does not measure. Ask for the rate of rejected looks and the rate of wrong-team suggestions, in writing, before accepting 10% as a reason to buy.
We would also treat engagement as a placeholder until it has a definition. If the metric is clicks on a suggested item, a 10% lift can come from placement alone; if it is revenue per session, it is a different purchase decision. The page gives no way to tell which one it is. For a store heading into a season peak, the calendar is a confounder the reader cannot rule out, and the honest reading of this release is that a large retailer got value from a working recommendation loop built in a week, with a customer-authored guardrail layer doing the quality work.
Sources
- How Fanatics and OpenAI streamline the fan experience, openai.com, published October 26, 2026: https://openai.com/index/fanatics-streamline-fan-experience
- The figures quoted above, read on that page on October 26, 2026: 10% higher engagement than the previous Shop the Look, roughly 30,000 MLB products in recommendation testing, under 24 hours for the first 50 products, MVP approved the same week
- Our earlier coverage of OpenAI build-speed claims: one-day repeat rollout and Block’s patch time