What a shopping agent sees on your store, and how to test it

The store an agent shops in

A shopping agent never sees your theme, your photography or your homepage film. It reads structured surfaces: the discovery files Shopify serves on every store, the catalogue through the storefront MCP interface, the policies and FAQ feed, and the structured data on your pages. Everything it knows about your store comes from those four, and every one of them is public.

That is the useful part. Because the surfaces are public, you can read your store the way an agent does, today, without installing anything, and the gap between what you tell shoppers and what an agent is being told is usually visible within the first ten minutes. This article is the test process we run at the start of every agentic readiness engagement, written so you can run it yourself.

The four surfaces, in the order an agent meets them

The four tests below take the surfaces in that order. For each one, note what you find, because the same reads repeated after the fixes are your before-and-after evidence.

Test 1: read your discovery files

Open yourstore.com/agents.md, /llms.txt and /llms-full.txt in a browser and read all three. On most stores all three match except for a single line identifying the path, which means the store serves Shopify's default and there is no brand content in the discovery layer. Note whether the policy links resolve and whether anything in the file misstates the store.

This test takes five minutes and requires nothing technical, and it is the one we recommend every merchant runs this week regardless of what else they do.

Test 2: search your catalogue the way an agent does

The storefront MCP endpoint speaks plain JSON over HTTP, so a developer, or any AI assistant with a code tool, can query it in minutes. The request calls the search_catalog tool with a natural-language query and buyer context, and the response is the product set an agent would work from, with the exact titles, descriptions and prices it would read.

Choose queries the way buyers phrase needs: a use case, an occasion, a constraint. For a country footwear store, "waterproof boots for a shooting day" tells you more than "boots", because it tests whether the catalogue's language matches the buyer's. Run your two or three most commercially important queries and read the responses closely.

• Are the right products first? If a use-case query surfaces the wrong range, the product data does not carry the use case, and no downstream layer can fix that.
• Do the descriptions give an agent anything to recommend with? A two-sentence description with no material, fit or use-case detail produces a recommendation with the same holes in it.
• Is anything commercially important missing entirely? An absent product line usually traces to catalogue structure or channel eligibility, and it is better found by you than by a customer's agent.

A worked catalogue read

This is what the assessment looks like in practice, for an illustrative premium skincare store queried with "gift set for sensitive skin under £60". The response carried four products, and reading them as an agent would produces four different verdicts.

One query, and the store's product-data quality is legible line by line: the piece that was written for this question wins it, the bestseller loses on missing information it possesses in real life, and nothing about the ranking is mysterious. Multiply by your top ten buyer questions and you have a catalogue work-list ordered by revenue.

What an agent-readable description contains

The moisturiser above fails as an agent read while being a perfectly good product page for a human, because the photography, reviews and brand context that persuade a person never reach the feed. The description field has to carry the load alone.

Written for humans only: "Our hero moisturiser, loved for years. Silky, weightless, and the finishing touch to any routine." Rewritten to carry its facts: "A daily moisturiser for normal to sensitive skin, fragrance-free and dermatologically tested, with squalane and ceramides. 50ml, suitable under make-up, and the partner product to the sensitive cleansing range." The second version is still brand voice, and it now answers skin type, size, formulation and adjacency, which are the fields an agent recommends on.

The pattern to apply across the catalogue: state what the product is, who and what it is for, the material or formulation facts a buyer would ask about, and what it pairs with, inside the first hundred words. Everything an agent cannot read from the description, it treats as unknown.

Test 3: ask the questions a shopper would ask

The same endpoint serves search_shop_policies_and_faqs, the tool an agent calls for returns, shipping, services and every other store question, drawing on the Knowledge Base feed. Query it with a written question set covering returns, exchanges, refund timing, delivery costs and times, collection, gifting, care, sizing and warranty, marking each answer accurate, wrong or missing.

Twenty to thirty questions is a full baseline, the reusable set is in our Knowledge Base guide, and the scoring matters more than the count. A missing answer is a gap an agent fills with a guess or a shrug, and a wrong answer is being relayed to buyers as fact until someone corrects the feed.

Test 4: watch the channel itself

Finally, test from the buyer's side. Ask ChatGPT, and any other platform your customers use, two or three category questions without your brand name: the best options in your category for a given need, budget or occasion. Note whether your brand appears, which competitors do, what language the platform uses about each, and whether the details it states are accurate.

Keep the queries identical each time you repeat the exercise, because this is the noisiest of the four tests and only consistent queries make movement readable. What you are measuring is presence and accuracy in the answer layer, and the earlier tests explain most of what you see here.

Test 5: check who you are blocking

One five-minute check sits outside the four surfaces and undoes all of them when it fails: robots.txt. Read yourstore.com/robots.txt and look for disallow rules against the AI crawlers, GPTBot, Google-Extended, PerplexityBot, anthropic-ai, Amazonbot, and Bytespider where a mainland Chinese customer base matters. Blocking rules usually date from a 2023-era decision, made when opting out of model training looked prudent and before the same crawlers became the route to being cited.

Shopify stores can edit the robots.txt.liquid template directly, so removing a stale block is a small theme change. Decide the policy deliberately either way, because a store blocking the crawlers while investing in AI visibility is paying for a channel it has barred at the door.

Testing the transaction path

Discovery and representation are most of the picture, and the last yard is whether an agent-driven purchase completes. The storefront interface carries cart tools alongside search, so a developer can walk the full journey: search the catalogue, build a cart, and confirm the handoff to checkout, which by protocol rule always requires the human buyer's approval.

For most merchants the practical check is simpler than a full agent build. Confirm in the agentic storefronts section of the admin which channels are live, place a test order from a channel you can access, and watch where the buyer lands: on mobile the purchase completes in an in-app browser, on desktop the platform links out to your own store, so your checkout experience is still the closer either way. Anything broken on that path, a forced account wall, a market mismatch, an expired discount banner, now fails in front of an agent-referred buyer at the moment of highest intent.

What these tests surface on live stores

Running this process across live stores, the findings repeat. Discovery files are the untouched default everywhere. Policy feeds misquote terms, answer location questions with raw internal IDs, and return nothing for a third to a half of a standard question set. Catalogue searches surface the right products with descriptions too thin to recommend from.

The structural findings are quieter and matter as much. On one multi-store brand, four of nine collection references inherited from planning documents were broken on the live stores, and nothing flagged it, because a request for a missing collection returns success with an empty products array. On another store the myshopify subdomain did not match the store name it appeared to belong to, and every query against the assumed store was silently answering for a different market. Verify the store's identity from currency and product URLs in a live response before trusting anything else the endpoint tells you.

The traps that produce false results

One date to hold: the cart tools on /api/mcp retire on 31 August 2026 as part of the UCP migration, and responses from the old endpoint already carry the deprecation notice inline. Audit tooling built this season should expect to update its endpoints next season, which is itself a lesson in how fast this layer is moving.

What testing will not tell you

The tests establish what agents can read, and they do not predict how often your brand gets recommended, because that rests on the brand's standing in the models' wider picture: authority, coverage, entity clarity, and signals well beyond the store. A clean feed and a rich catalogue make the store safe to recommend and accurate when represented, and the being-chosen work is the broader discovery-ready programme.

Treat the results as a snapshot, dated and repeatable, and resist the urge to fix the findings in the order they annoyed you. Catalogue and policy corrections come first because every other layer restates them, then the Knowledge Base, then the discovery files, and then the channel watch runs on a calendar rather than on adrenaline.

Who runs what, and how often

The whole cycle is under three hours a quarter once the question sets exist, and it is the difference between knowing what agents say about the store and assuming it. The first run is the one that produces surprises, so run it before your busiest season, when there is still time to act on what it finds.

Questions we hear about agent testing