A 90-Day Plan to Make Your Store Callable by AI Agents

Everything in the Fall '26 release that matters for GEO can be turned into roughly a quarter of work. Not a transformation programme — a sequence of mostly cheap, mostly one-off tasks, in an order that matters more than the tasks themselves.

The order is bottom-up: prove you can be called before you polish what the call returns.

Key takeaways

• Weeks 1–2: prove the plumbing. Six probes, plus catalog eligibility. Acceptance is binary.
• Weeks 3–6: make the data readable. Taxonomy, the three supported attributes, stock hygiene, extractable copy on the top 50 SKUs.
• Weeks 7–10: settlement and trust. Knowledge base published, payment handlers verified, a deliberate decision on agent checkout.
• Weeks 11–13: measure and scale. Server-side attribution, a version boundary in reporting, an agent-channel line, then evaluate paid placement.
• Six roles own distinct pieces. If this sits entirely with marketing, it will not finish.

Entry condition: are you even eligible?

Before week one, confirm Shopify Catalog eligibility. The bar is low but absolute:

• Compliant with Shopify's terms and acceptable use policy; a paid plan; the store not password-protected.
• Every product has a title, at least one image, and a price above zero — free products are excluded.
• Products published to Online Store, Hydrogen or a headless channel.
• Terms of service, privacy policy and refund policy complete under Settings → Policies, with the agentic storefronts supplementary terms accepted.
• Channel gates on top: ChatGPT and Microsoft Copilot cover merchants selling to US buyers regardless of where the business is based; AI Mode and Gemini are currently limited to a subset of US-based merchants.

Fall '26 added an eligibility check in the admin, so this is a lookup rather than a guess.

Weeks 1–2 · Prove the plumbing

Run the six probes from the readiness audit: both agent files returning 200, eight or more UCP capabilities, payment_handlers not empty, permalink on your own domain, roughly fifteen tools, WebMCP markers above zero.

Acceptance is all six green. Not five. There is no partial credit at this layer — a failed tool call returns nothing, not less.

Most failures here are configuration, and most are found on headless builds. Budget a few engineering days, not a sprint.

Weeks 3–6 · Make the data readable

Now the product work pays, because there is something to pay into.

• Standard taxonomy on every core SKU. Recall runs on broad category terms, so this decides whether you are in the candidate set at all.
• Color, Size and Target gender filled correctly. These are the only three attribute names the cross-store catalog accepts; everything else is ignored.
• Stock hygiene. Search returns only purchasable items. Out of stock equals non-existent.
• Rewrite the top 50 products as extractable fact. This is the highest-leverage content work of the quarter — see product copy for extraction.
• GTIN wherever one exists. Lookup accepts GTIN and UPC, which makes it the cross-source identity key.

Acceptance: query the catalog for your own core category terms and confirm you come back.

Weeks 7–10 · Settlement and trust

• Publish the knowledge base. Ten fact categories, 1,000-fact ceiling, and nothing counts until published. Why to do it before the protocol settles.
• Verify payment_handlers and native_checkout eligibility. The latter defaults to false per variant, and in-chat closability is a reason for an agent to prefer you.
• Decide on Google's agentic checkout deliberately. It was enrolled by default; inheriting a default is not a decision. Read what breaks in attribution before you choose.
• Qualify for Shop Promise. Official wording: products with the badge rank higher. It is the only ranking lever here that is pure operations and not an auction.

Weeks 11–13 · Measure and scale

• Move attribution server-side. GA4 Measurement Protocol, Meta CAPI, Google Ads enhanced conversions, all sourced from the order store.
• Draw a version boundary in reporting and rebuild the baseline. Sessions and conversions both changed in the same quarter; comparisons across that line are meaningless.
• Make agent-sourced orders their own line. If they are not separately countable, you will spend next year arguing about whether any of this worked.
• Then, and only then, evaluate paid. Promoted placement blends into organic ranking, so capture the organic baseline first or you will never know what the budget bought.

Who owns what

Two honest caveats

Sampling. The measurements behind this plan come from a small number of live stores on 29 October 2026. They demonstrate that these failure modes occur; they do not tell you how common they are. Run the probes against your own domain.

Proportion. Agent-sourced volume is still far smaller than organic search and paid for most brands. This plan argues for fixing the foundation, not moving the budget. The foundation is cheap — largely one-off configuration and data cleanup — and that cheapness is exactly why waiting for more data is the wrong call.

FAQ

Can a small team do this? Yes. Weeks 1–2 are a few engineering days. Weeks 3–6 are the real work and scale with SKU count — capping it at the top 50 keeps it to one person for a month.

What if we are mid-replatform? Add the four agent-visibility checks to the migration acceptance criteria now. It costs ten minutes per build and prevents the most expensive version of this problem — see the headless audit.

How do we know it worked? Three indicators, in order: all six probes green and staying green; your own products recalled by core category terms in a tool call; agent-sourced orders visible as a separate reporting line with a trend. The first is binary, the second is diagnostic, the third is the business case.