After Fall '26, the surface a Shopify store exposes to AI agents can be described in six layers. Top to bottom is the order an agent touches them. Bottom to top is the order you should build them.
Getting that inversion right is most of the value of this model, because almost every team we see starts at the wrong end.
Key takeaways
- Six layers: channel, skill, service, protocol, data, settlement.
- Each has its own failure mode and its own owner. Treating them as one project is why agentic work stalls.
- Layer five, data, is the only layer almost entirely under your control.
- Layer four and layer six are platform-hosted — which makes them easy to assume are fine when they are not.
- Build bottom-up: prove you can be called before you polish what the call returns.
The six layers
Layer | Key assets | What you control | Most common failure |
|---|---|---|---|
1 · Channel | ChatGPT, Copilot, Gemini and AI Mode, the Shop app | Whether you are eligible, and in which markets | Assuming eligibility equals volume, and ignoring each channel's own gate |
2 · Skill | The official Shop skill, your | Not the file itself — only what it cites | Writing an llms.txt by hand, when on Shopify it mirrors agents.md |
3 · Service | Global Catalog, Storefront Catalog, cart/checkout/order, Ask, WebMCP | Whether you are indexed, whether tools are registered | A headless storefront missing the whole layer, unnoticed |
4 · Protocol |
| Mostly platform-hosted — your job is to confirm it is up | Capabilities on a custom domain differing from the main store |
5 · Data | Title, description, media, options, taxonomy, price, list price, stock, facts | Almost all of it | Spending the budget on PDP visuals agents cannot see |
6 · Settlement | Shop Pay, | Payment methods, markets, agentic terms |
|
The counter-intuitive priority
The instinct is to start at layer five. It is the layer marketers own, the work is familiar, and it feels productive: rewrite titles, enrich descriptions, fix taxonomy.
But if layer four or layer six is broken, better layer-five work only raises the quality of an interface that will never be called. You cannot improve your way out of a 404.
Build bottom-up:
- Layer six first. Is
payment_handlerspopulated? Does the permalink point at your own domain? If a sale cannot close, nothing above matters. - Layer four next. Does
/.well-known/ucpreturn eight or more capabilities? This is usually fine, and it takes one command to be sure. - Layer three. Does
tools/listreturn around fifteen tools, and does the homepage register WebMCP markers? - Layer two. Do
/agents.mdand/llms.txtreturn 200, and are the policies they reference complete? - Layer five. Now the product data work pays, because there is something to pay into.
- Layer one. Channel eligibility and market expansion, once the rest holds.
The ten-minute readiness audit covers steps one to four with six commands.
Each layer's real owner
This is where the model earns its keep organisationally. These are not all marketing's job, and pretending otherwise is why the work stalls.
Layer | Owner | First action |
|---|---|---|
1 · Channel | Growth | Confirm eligibility and your actual switch state per channel |
2 · Skill | Legal and support | Complete the three policy pages; accept the agentic terms |
3 · Service | Engineering | Verify WebMCP registration and the tool list |
4 · Protocol | Engineering | One curl, quarterly, and after every release |
5 · Data | Merchandising and content | Taxonomy, Color/Size/Target gender, extractable copy |
6 · Settlement | Payments and finance | Payment methods, markets, |
A brand with a strong layer five and a broken layer six is the most common pattern we see, and also the most wasteful — the investment is already made and the return is zero.
Where the layers break in practice
Headless builds break layer two and layer three together. Both agent files 404 and WebMCP registers nothing, while layer four still returns 200 — which makes it look like the protocol is fine. The full comparison is here.
Custom domains break layer four and six quietly. Capabilities and permalink endpoints can differ from the main store, so the permalink points back at a myshopify domain and nobody notices because the storefront works perfectly for humans.
Marketing breaks layer five by optimising the wrong fields. The product object an agent receives has 14 top-level fields, and reviews, video and 3D are not among them.
FAQ
Do I need all six layers working to get any value? You need four, five and six. Layer three is largely automatic on Liquid. Layer two depends on layer six's policy prerequisites. Layer one is where growth happens once the rest holds.
How often should this be audited? Quarterly, plus after every theme release, migration, or payment and market change. Protocol versions move independently of your release schedule.
Is this model Shopify-specific? The layer structure generalises to any commerce stack. What is Shopify-specific is how much of layers three and four you get by default — on a self-built stack those are engineering budget, which is the sharpest competitive point in the whole release.