Shopify

Shopify provides a commerce platform for businesses selling online and in stores. Its AI work spans Sidekick, a merchant business assistant; SimGym, which simulates storefront visits; and Tangle, an open-source platform for reproducible ML experiments. Alongside these products, Shopify made AI use an employee expectation and partnered with Liquid AI on models for search and recommendations.[15][4][3][2][8]

Shopify introduced its vision for Sidekick in summer 2023 and shipped a new architecture in 2025 to improve its performance. The Winter ’26 Edition, announced on December 10, 2025, expanded it into more complex tasks, automations, theme editing and email editing. Sidekick Pulse added proactive recommendations based on both store-specific and Shopify-wide signals, with assistance in carrying out recommended actions.[4]

Sidekick also gained custom app generation inside the Shopify admin, using Polaris UI and the Admin GraphQL API. Merchants could visually iterate on an app, test its changes and install it. Shopify separately released Sidekick App Extensions in developer preview, allowing third-party apps to expose data and functionality in merchant conversations—for example, answering questions about an app’s context or navigating to its relevant screens.[4]

In an April 2026 interview, CTO Mikhail Parakhin identified Andrew McNamara as the head of Sidekick and Pulse. Parakhin said McNamara had also been the first development lead for Sydney, Microsoft’s earlier Bing chatbot.[1]

In April 2025, CEO Tobi Lütke declared that “reflexive AI usage is now a baseline expectation at Shopify.” Distinguished Engineer Daniel Beauchamp’s October 2025 account reported unlimited access to top AI models for every team. Quick, an internal sandbox for prototype sites launched in mid-2025, had grown to more than 5,000 sites, according to Beauchamp. He also described Scout, an AI tool aggregating hundreds of millions of merchant feedback items.[2]

By April 2026, Parakhin described employee token funding as effectively unlimited, with employees choosing their tools. Shopify discouraged using models below Opus 4.6 rather than setting an upper limit on model capability. He said CLI-based coding tools were growing faster internally than tools requiring an IDE.[1]

Parakhin argued that code volume had made review and deployment the limiting steps. Even if models produced fewer bugs per unit of code, generating much more code could let more bugs reach production. Shopify used its own AI PR review flow: strong models took turns critiquing and revising work, rather than operating as a large parallel swarm. In his explanation, spending longer on review could save time later by avoiding failed tests, removal of offending PRs and deployment rollbacks. Shopify also used stacked PRs and worked with Graphite.[1]

Shopify open-sourced Tangle on December 5, 2025. The Search & Discovery team developed it from Cloud Pipeline Studio, a tool Alexey Volkov had built before joining Shopify. Tangle addressed repeated data preparation, lost experiment history and the difficulty of moving notebook-based work into production. Shopify reported more than a year of accumulated compute-time savings from using it.[3]

Tangle combines a visual directed acyclic graph with reusable components defined in YAML. Components wrap containerized command-line programs, allowing teams to connect code written in different languages without rewriting it for a shared framework. Programs exchange files; Tangle manages storage transfers and runtime file locations. Components are designed to behave as deterministic, side-effect-free functions, providing the contract needed for caching and reuse.[3]

Its content-based caching checks output hashes: when an upstream change produces identical output, downstream work can reuse cached results. Shared preprocessing can run once for several experiments, and the cache can reuse both completed and still-running executions. Preserved graphs, logs and immutable intermediate artifacts let colleagues clone a run, change a component and compare results. Parakhin emphasized that the same experiment could run in production without being ported to another system.[3][1]

Parakhin described Tangent as an internal agent operating above Tangle. It modifies and reruns experiments toward a measured objective, either recombining existing components or inspecting code and creating new ones. He described applications in theme latency, search optimization, prompt compression and storage reduction, and said its users extended beyond ML engineers to product managers bringing domain knowledge.[1]

Parakhin’s qualification was that automatic research worked well on overlooked optimizations and standard practices, but struggled with entirely new approaches. In one personal experiment on a problem already optimized for years, he recalled running more than 400 experiments over several weeks and finding only one successful improvement. The benefit, in his account, was obtaining that improvement without conducting all those experiments himself.[1]

Shopify announced SimGym in Research Preview in December 2025; its app listing records a December 8 launch. SimGym uses AI shoppers with human-like personas to explore storefronts and provide signals about navigation and add-to-cart behavior before changes go live. It differs from Rollouts, announced in early access that month, which stages changes and can allocate real traffic between variants for A/B testing.[6][4]

Parakhin said SimGym relied on historical commerce behavior to calibrate its shoppers, rather than simply prompting agents in isolation. The team spent almost a year optimizing simulations, with an internal goal of 0.7 correlation with add-to-cart events; he did not report that figure as an achieved result. He also said a store’s own customer history improved forecasts by helping agents reproduce its particular customer distribution.[1]

SimGym initially compared two storefront variants. Parakhin recalled that the team then recognized many smaller merchants had only one configuration and needed suggestions for what to change. By April 2026, he said Shopify had released a version that examined a single storefront and recommended changes intended to improve conversions.[1]

The simulations use visual browser interaction so shoppers encounter presentation and friction resembling what people see. Parakhin described the expense of multimodal models and browser farms, and optimization work with Fireworks, CentML and NVIDIA. Serving costs, model distillation and cheaper browser execution were major concerns as usage increased.[1]

Liquid AI announced a multi-year licensing and R&D partnership with Shopify in November 2025. It followed Shopify’s participation in Liquid AI’s US$250 million Series A in December 2024; Shopify’s individual investment amount was not disclosed. Liquid AI reported a first production search-model deployment below 20 milliseconds and a jointly developed HSTU generative recommender that improved recommendation conversions in controlled testing, without publishing a numerical uplift. Exact agreement terms were undisclosed.[8]

Parakhin described two distinct Liquid model workloads in April 2026: a roughly 300-million-parameter model for query understanding at about 30 milliseconds end to end, and larger task-distilled models for offline catalog processing and Sidekick Pulse. The search work involved Liquid AI, CentML and NVIDIA. Batch catalog tasks emphasized throughput—assigning taxonomy, normalizing attributes and identifying products sold by multiple merchants—while interactive search demanded low latency. These were specialized deployments, rather than a claim that Liquid models matched frontier models across tasks.[1]

Shopify’s December 2025 Edition opened Catalog API access through its developer dashboard and introduced Agentic Storefronts to syndicate merchant products into AI shopping channels. Shopify Catalog infers categories and attributes across billions of products and clusters identical products to reduce duplicate search results. The developer offering combined product discovery with a universal cart API and Checkout Kit for embedding a merchant’s checkout.[4]

Shopify also co-developed the Universal Commerce Protocol with Google. UCP is an open standard in which merchants and agents declare their supported commerce capabilities and negotiate how to interact. Its design includes extensible commerce workflows, negotiated payment handlers and multiple transports, including REST, GraphQL and MCP. Parakhin discussed Shopify’s UCP and catalog work in April 2026, including runtime product search, bulk lookups and identity linking.[9][1]

The storefronts and developer tools behind these AI products predate Shopify’s recent AI expansion. Shopify’s origin lay in Lütke’s Snowdevil snowboard store and the commerce software he built for himself. In a 2024 retrospective, he recalled payment-setup friction that included sending notarized passport copies to obtain credit-card API keys. His explanation for starting Shopify was: “The software I built for myself is good. Maybe other people want it too.”[10]

On June 2, 2009, Shopify officially launched its API Platform and App Store, reporting more than 5,000 merchants. Developers could create and sell applications for individual store needs. Lütke explained that installing selected features would give merchants customization without making the core software cumbersome. Initial apps included invoice printing, customer reviews and digital-goods sales.[11]

Shopify POS entered brick-and-mortar stores in 2013. It evolved from serving casual sellers and single-location retailers toward multiple locations and unified online and offline features, such as in-store pickup and centralized customer profiles. In December 2025, Shopify announced POS Hub, which actively manages peripheral connections and supports wired scanners, card readers and printers.[4]

On May 4, 2023, Lütke announced that Shopify would become about 20% smaller and that Flexport would buy Shopify Logistics. His signed memo framed logistics as a worthwhile “side quest” that should become another company’s main focus, while Shopify concentrated on commerce software. He explicitly connected that focus to emerging AI capabilities and the possibility of a copilot for entrepreneurship.[13]

Shopify completed the sale of the majority of its former logistics business to Flexport on June 6, 2023. It received stock representing an additional 13% equity interest in Flexport, beyond its existing holding.[14]

On May 27, 2015, Shopify announced the closing of its initial public offering, including exercise of the underwriters’ over-allotment option, raising US$150,535,000 in aggregate gross proceeds.[12]

For the quarter ended June 30, 2026, Shopify reported US$3.583 billion in revenue, up 34% from US$2.680 billion a year earlier. Subscription solutions contributed US$802 million and merchant solutions US$2.781 billion. The August 5 release described millions of businesses using Shopify in more than 175 countries.[15]

Quarterly gross merchandise volume was US$115.567 billion, compared with US$87.837 billion in Q2 2025. GMV measures the value of platform-facilitated orders, including specified revenue-sharing apps and channels, net of refunds and inclusive of shipping, handling, duty and value-added taxes. It is distinct from Shopify’s own revenue.[15]