Shopify Magic, Sidekick, and Inbox AI are real now. So are OpenAI, Anthropic Claude, and Google Gemini integrations into Shopify Plus. After eight production AI builds we have a clearer view of what each is for — and where it earns its token budget.
Shopify shipped Magic, Sidekick, and Inbox AI across 2024–25. By 2026 they are everywhere — built into product creation, theme content, the admin assistant, and chat suggestions. They are also, in many cases, not enough.
This is what we have learned from eight Shopify AI integrations shipped between 2024 and 2026. Where Shopify Magic is the right answer, where you need a custom OpenAI or Anthropic Claude integration, and where AI is the wrong answer entirely.
Shopify Magic is Shopify's native AI feature suite. It is built into the admin and ships free with every plan. It does five things well:
For stores under ~500 SKUs with simple content needs, Magic is genuinely enough. The configuration is a one-week sprint — set the brand voice prompt, train a couple of internal users, document the workflow, hand over.
For stores above 1,000 SKUs, Magic alone is not enough. Not because it is bad — because the workflow ergonomics fall apart at scale.
Sidekick is Shopify's AI assistant inside the admin — an embedded chat interface that can do tasks, query reports, and walk merchants through configuration. It is included with most Shopify plans.
Out of the box, Sidekick handles standard admin tasks: "show me top sellers this month", "create a discount for this collection", "check inventory in this warehouse". For a 1-store, single-region merchant, no configuration needed.
Configuration earns its keep at three points:
When you graduate beyond Magic and Sidekick to custom integrations, the model choice matters. Eight builds in, this is our current rule of thumb:
Best for generative content, structured-output tasks, image generation. Largest ecosystem of tooling. Default choice for product description rollouts and PDP content generation.
Best for long-context work, complex customer-service triage, and where safe-refusal behaviour matters. Wins on catalog-wide reasoning (full SKU sets in context) and at customer-care use cases.
Best for multimodal (image + text) and where pricing matters. Wins on visual product matching, image-to-product search, and high-volume cheap inference.
Pick by job, not by hype. We benchmark on the client's actual data before committing — and on three of eight builds we ended up running two providers (OpenAI for one feature, Claude for another) because no single model wins everywhere.
Traditional Shopify search (Algolia, Klevu, Searchanise on default config) is keyword-matched with synonyms and filters. AI search uses embeddings — searches by intent, not by exact word match. "Warm winter coat" returns the right products even if no product mentions "warm" in the title.
The lift on stores doing >£100k/month is consistent: 15–30% increase in search-to-conversion. Below that traffic threshold, the improvement does not pay back the implementation cost.
Implementations we have shipped:
Rules-based recommendation apps (LimeSpot, Rebuy, Nosto) are excellent at collaborative-filtering — "people who bought X also bought Y" patterns from your purchase history. They cap at the patterns in your existing data.
LLM-driven recommendations handle three things rules-based cannot:
In practice we run both: rules-based for the standard PDP slot, LLM-driven for the high-AOV upsell layer. Different jobs.
Conversational commerce — AI chatbots that sell, not just answer questions — is the most-hyped, least-shipped Shopify AI use case. We have built two production conversational flows. Both took 6–8 weeks. Both lifted AOV measurably. Neither replaced human service entirely.
Where conversational commerce works:
Where it does not yet work:
Token cost is the variable nobody talks about until they ship. Eight builds in, our typical run-rates:
Every AI integration we ship has a token-cost dashboard from day one. Cost per request, cost per conversion, cost vs uplift. AI without cost discipline is a way to lose money confidently.
The honest list:
AI eCommerce is a tool, not a strategy. The stores winning with it picked specific use cases, measured the lift, and shipped only what earned its token budget. That is the work.
