NEWS

Shopify's new search and recommendation model is trained to surface stores a shopper has never bought from

Shopify now runs one generative model behind search and recommendations. It lifted orders from the Shop home feed by 7.95%, and it places each product by what its listing says.

A canvas backpack hanging on a door
Photo by Ruben Valenzuela on Unsplash

On November 11, 2026, Shopify's engineering team described a single generative model that now proposes products for both search and recommendations, and it is running in production. Shopify's example is two people who search for "backpack". One has been comparing waterproof laptop bags and the other has been looking at larger packs for hiking in the Rockies, and the model uses what each has explored and bought to show them different products. Shopify trained it on search, browsing and purchase activity across millions of participating merchants. The stated goal is for each buyer to find relevant products from merchants they know and from merchants they have never visited.

Shopify reports what happened when it added the model to the home feed of its Shop app. Orders rose 7.95%, the number of buyers who purchased rose 8.60%, and the conversion rate rose 9.85%. In search, the model raised Shopify's relevance score for the top ten results from 30.6 to 32.7, and for signed-in shoppers a quarter of queries returned better results.

Conversion rate+9.85%Buyers who purchased+8.60%Orders+7.95%Relative lift on the Shop home feed after adding the generative model. ShopifyEngineering, Nov 11, 2026

The way the model identifies a product is the part a merchant can act on. Shopify's catalog holds several billion products, too many for a model to learn one by one. So Shopify turns each product's content into a short code of three numbers, and products with similar content share the start of that code. Shopify's own example is a group of vests and gilets that share a first number, with padded gilets sharing the second and gray padded ones the third. A product with few clicks or sales still benefits, because the model learns about it from the neighbors that share its code.

That makes the listing the thing that places a product. A new item with no sales history is filed next to similar items by its description and attributes, and it is offered to the shoppers those items suit.

Shopify also trained the model to favor discovery. Shopping histories are full of repeat visits, so a model trained on them leans toward what a buyer already knows. Shopify gave extra weight to three events: a buyer's first purchase of a product, first purchase from a merchant and first purchase in a category. In its tests, that raised the share of purchases the model anticipated among products the buyer had never interacted with by roughly 10% to 15%.

The model's job is to nominate candidates, and later stages rank them. Shopify notes that a relevant product missed at this first step cannot be recovered by a later ranker. Its next plans are to have the model read product images for shape, texture, pattern and style, and to extend the same model to ads.

If you sell on Shopify, open the product you added most recently and read its title, description and attributes as if you had never seen the item. State what it is, what it is made of and who it is for, since that text is what files it beside the right neighbors.

Common questions

What did Shopify announce on November 11, 2026?

Shopify's engineering team described a shared generative model, trained on both search and recommendation data, that now generates product candidates for both in production.

What results did Shopify report?

On the Shop home feed, orders rose 7.95%, buyers who purchased rose 8.60% and the conversion rate rose 9.85%. In search, Shopify's relevance score for the top ten results rose from 30.6 to 32.7.

How does the model decide where a product belongs?

Shopify converts each product's content into a three-part code. Products with similar content share parts of the code, so the model can learn about a product with little history from similar products.

Does the model favor stores a shopper already knows?

Shopify trained it to do the opposite for recommendations. It gave extra weight to a buyer's first purchase of a product, from a merchant or in a category, which improved its recall of purchases of unfamiliar products by roughly 10% to 15%.

Sources

Loading Lantern…

Use your browser’s Reload button to try again.