---
type: "post"
id: "shopify-s-generative-discovery-model-fd2f6f"
slug: "shopify-s-generative-discovery-model-fd2f6f"
title: "Shopify’s generative discovery model"
description: "Shopify says a shared generative model for search and recommendations is now in production, with reported gains in Shop home-feed outcomes and search relevance."
retrieval_nugget: "Shopify says a shared generative model for search and recommendations is now in production, with reported gains in Shop home-feed outcomes and search relevance. The case matters to AI builders because it shows how a shared product representation and joint training can support multiple retrieval tasks at catalog scale."
published_at: "2026-11-12"
updated_at: "2026-11-12"
record_date: "2026-11-12"
date_kind: "published_at"
topics: ["models","retrieval"]
entities: ["Shopify","Shopify Shop","RQ-KMeans","RQ-VAE"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "Shopify describes a shared generative model that personalizes product search and recommendations using shoppers’ browsing and purchase context."
prior_refs: []
source_url: "https://shopify.engineering/generative-discovery"
source_urls: ["https://shopify.engineering/generative-discovery"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:shopify-s-generative-discovery-model-fd2f6f"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/shopify-s-generative-discovery-model-fd2f6f/","markdown":"https://newruntime.com/posts/shopify-s-generative-discovery-model-fd2f6f.md","json":"https://newruntime.com/posts/shopify-s-generative-discovery-model-fd2f6f.json"}
markdown_url: "https://newruntime.com/posts/shopify-s-generative-discovery-model-fd2f6f.md"
---

# Shopify’s generative discovery model

## Retrieval answer

Shopify says a shared generative model for search and recommendations is now in production, with reported gains in Shop home-feed outcomes and search relevance. The case matters to AI builders because it shows how a shared product representation and joint training can support multiple retrieval tasks at catalog scale.

## Summary
Shopify says a shared generative model for search and recommendations is now in production, with reported gains in Shop home-feed outcomes and search relevance. The case matters to AI builders because it shows how a shared product representation and joint training can support multiple retrieval tasks at catalog scale.

## What happened
Shopify’s engineering announcement, published November 11, 2026, describes a generative discovery model trained jointly on the company’s search and recommendation data. The model generates product candidates for both experiences. Shopify says it is deployed in production and that it improved offline and online measures. The post points readers to a detailed technical report for the full methodology and experimental results.

The system starts from an open-weight pretrained language model, then adapts it using commerce data Shopify says is available for this purpose: its product catalog and eligible buyer interactions across search and recommendations. Shopify’s catalog contains several billion products, too many for a useful model vocabulary made from independent raw product IDs. Instead, product content is encoded as an embedding and converted into a short semantic ID. Shopify evaluated RQ-VAE and RQ-KMeans and selected RQ-KMeans based on better downstream performance at the same codebook shape. Its selected configuration uses three codebooks of 8,192 entries, yielding a three-token ID per product. Multiple products can share a full code; at serving time, Shopify explored collision-resolution approaches, including lightweight ranking using popularity.

For recommendations on the Shop home feed, Shopify reports relative online lifts of 7.95% in orders, 8.60% in converters, and 9.85% in conversion rate. Offline, the model adds 2.65 percentage points of Incremental Pool Recall to the recommendation production system and is described as the company’s strongest standalone candidate generator. For search, Shopify reports that adding the model raises combined OneDCG@10 from 30.6 to 32.7, a 2.1-point gain; for signed-in users in the personalized-only configuration, it rises from 33.0 to 35.7, a 2.7-point gain. Shopify says 25% of queries improve on net in top-10 relevance. OneDCG@10 is described as a variant of NDCG that normalizes against ten perfect results for every query.

The post also describes training choices. Shopify combines search and recommendation learning and reports that removing one data stream can reduce the other task’s measures. To promote discovery of products buyers have not previously interacted with, it upweights first-purchase examples and compresses histories by dropping stale clicks while retaining purchases. In an offline experiment, Shopify says this combined history-compression and discovery recipe improved discovery recall by roughly 10–15% relative at cutoffs of 10 and 50, while overall recall rose about 2% relative. The announcement’s text further describes an LLM relevance judge as a reward signal for search reinforcement learning.

## What the archive adds
[New Runtime’s earlier coverage of Shopify’s task-expert loop](https://newruntime.com/posts/how-shopify-turns-open-models-into-task-experts-edc9e2/) appeared on October 22, 2026, and treated a narrow agent workload as the unit of optimization. This is a new fact on a known Shopify story: the current announcement describes a shared discovery model in production and reports search and recommendation outcomes, rather than a restatement of the earlier task-expert coverage.

## Trend signal
This supports [agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work) only weakly: Shopify reports buyer and retrieval outcomes rather than token costs or verified agent-task costs. More directly, the case suggests that production model work is judged by task-level outcomes, but it does not establish an agent-economics result. This does not yet fit a pattern New Runtime tracks.

## Why now
Shopify says the model is already deployed and frames its work around serving recommendations and search across a catalog of billions within the latency budget of live requests. The announcement is therefore about a production system, not only a proposed architecture; it does not provide a launch date beyond publication or a rollout schedule.

## Why it matters
The useful engineering lesson is not simply that one model can serve two features. Shopify describes a common representation for products, shared training signals, and task-specific refinements intended to make search and recommendations help each other. That design can reduce the need to treat every catalog item as an isolated label, while semantic-code sharing gives sparse products a way to benefit from related examples.

For teams, the reported numbers are reasons to inspect the measurement design, not targets to assume. The search table distinguishes a production system with and without the model from the model alone, and the recommendation figures combine online and offline metrics. A team considering shared retrieval should test both task-specific and joint performance, define discovery separately from repeat engagement, and make serving-time handling of code collisions explicit. The architecture is potentially reusable; Shopify’s commercial outcomes and data access are not automatically reproducible elsewhere.

## Limits and uncertainty
The reported improvements are Shopify’s own results. The announcement excerpt does not include experiment duration, sample sizes, confidence intervals, or the detailed technical report’s complete methodology, so the strength and portability of the lifts cannot be independently assessed here. The model’s precise serving latency, infrastructure cost, open-weight model identity, and availability to outside developers are also not established in this account. Shopify says eligible buyer interactions are used for this purpose, but the announcement does not detail eligibility criteria or the governance controls around those data. The figures should be read as company-reported results, not an independent evaluation.

## Skill to build now
If your team has both search and recommendation, build an offline ablation that removes each task’s training stream in turn and measures both tasks. Shopify’s approach also suggests separating discovery from retargeting: measure purchases of items with no prior buyer interaction, then test whether first-purchase examples and history compression improve that metric without degrading overall retrieval.

## Reader takeaway
Shopify’s case is a concrete example of shared generative retrieval in production: one product vocabulary and joint training support both search and recommendation, with company-reported gains in each. The reusable lesson is to test cross-task benefit and discovery quality explicitly. The reported lifts are not a promise that the same model recipe or results will transfer to another catalog.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- topic hub: [Retrieval](https://newruntime.com/topics/retrieval.md): Explore the retrieval topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)