Audit one product, not a store. Write down the requirements its shoppers actually state. For each one, record what the received record says and in which field, then mark it: supported, ruled out, words only, two values, or unknown.
Then make one decision. Pick the single fix that matters most, give it an owner, name the document that proves it, and set a date to check it landed. That’s the whole output: one sourced change for this week. It isn’t a readiness grade, a missing fact stays unknown, and nothing here tells you whether an assistant will pick the product. The worksheet is our proposal. Below it’s filled in for a real catalog listing.
The worksheet in five parts
| Part | What you fill in | Done when |
|---|---|---|
| A. What you audited | Product, variant, route, request, buyer context, date | Someone else could pull the same record |
| B. Requirements | Each thing the shopper needs, its source, and what would count as support | Every threshold you chose is written down |
| C. Received evidence | What each field of the received record says about each requirement | Every cell is a quote with its field named, or “nothing” |
| D. States | Supported, ruled out, words only, two values or unknown | No cell is left as a guess |
| E. The fix | One change, an owner, the proof needed and a recheck date | One person has one job with one source |
Parts B to D are the coverage method from our constraint coverage guide, applied to a single listing. What’s new here is part E, which turns a grid of observations into one assigned piece of work. If you’ve already run the five-stage ACO plan, this worksheet is what you produce at its record and requirement stages.
Part A: pin down what you audited
Write the header first. Without it, a second person auditing “the same product” may be looking at a different variant, route or day.
| Field | What to write | Why |
|---|---|---|
| Product and variant | The variant you read, with its options | Facts and ratings can differ by variant |
| Route | Where the record came from: Shopify Catalog, a Google feed, your crawled page | Different assistants receive different records |
| Request | The exact words searched, if you searched | Results belong to a request, not a product |
| Buyer context | Country, signed-in buyer, or “none set” | Shopify’s catalog accepts country, region, language and currency |
| Date and evidence link | When the record was read and where the saved copy is | So you can compare it after the fix |
Before Part B, confirm the product can reach the route at all. Shopify’s Catalog requirements include a title, at least one image and a price above zero. The store must be on the Starter plan or higher and out of private mode. There is one exception: a store on the Agentic plan that hasn’t published its online store to a paid plan can still be included through agentic storefronts. If the product is ineligible, stop. No copy fix overrides exclusion.
Part B: write down the requirements and where they came from
Use the words your customers use: support emails, reviews, return reasons, the question you answer most often. Split each one into requirements, one per row, and write down what would count as support.
Two translations always need a decision from you. “Large room” needs an area, and “quiet” needs a decibel figure with a setting and a distance. Write the thresholds down. Different thresholds give a different audit, and the intent panel guide shows how to build a reusable list of these requirements from real customer evidence.
Add the offer basics to every audit, whatever the shopper asked: price, availability, ratings at product and variant level, and the shipping and refund policies. A product can answer every requirement and still lose the sale to a stale price.
Part C: read the record that was received, not your admin
Your product page isn’t what an agent gets. It gets a record, and some of that record may not be your words. Shopify marks the catalog description and its generated spec, feature and selling-point lists as Inferred: generated or enriched by Shopify, possibly absent or inaccurate, and to be treated as discovery signals rather than merchant-authored facts (Global Catalog MCP, “Inferred response fields”, checked October 27, 2026).
There are three documented ways to see a received record:
- Shopify’s search preview. Shopify says the preview in Sales channels > Agentic shows the raw outputs of Shopify Catalog search, and that AI channels often re-rank results, so it’s a directional signal (Managing agentic storefronts). We haven’t seen this screen ourselves.
- A catalog lookup by a developer. Shopify’s Global Catalog documentation describes a lookup that resolves known product or variant identifiers to their current catalog data.
- Your own product page, as the copy you control. It’s the source you’ll edit, not the record that was received.
Keep one column per field. “The record says 26 dB” is not the same finding as “a list Shopify generated says 26 dB”.
Parts D and E: mark the states, then choose one fix
The five states are the ones in our coverage guide: supported, ruled out, words only, two values and unknown. Unknown means the record doesn’t say. It never means zero, and it never means the product fails.
To choose the fix, work down this order and stop at the first row that applies:
| Order | What you found | The fix | Why it comes here |
|---|---|---|---|
| 1 | A wrong fact, or two values for the same requirement | Settle it at the source, then correct every copy | A wrong fact misleads the shopper |
| 2 | Words only on a stated requirement, where you hold or can get the source | Replace the adjective with the value, unit and condition | Cheapest fix with the clearest gain in checkability |
| 3 | Unknown on a stated requirement | Publish the value from a real document, or measure it. If it’s a limit, publish the limit | It may turn into a true “no”, and that’s useful |
| 4 | A number published without its condition | Add the setting, distance or method | It makes your number comparable |
| — | A true “no” | Nothing. Leave it | It sends the wrong shopper elsewhere |
Within a row, prefer the requirement your customers state most often, then the one you already hold a document for. Generated text is not a document. If Shopify’s list shows a value your page doesn’t, that’s a lead to trace, not proof to copy. Our claim-record guide ranks the evidence types.
Then write the fix as a card: the change, the owner, the proof and the recheck date. Only one fix goes to the top. Queue the rest underneath.
A worked audit: one humidifier record
On October 26, 2026 we searched Shopify’s agent catalog for “humidifier for dry winter air large room quiet top fill” and saved the top 10 results. We wrote that search to resemble a shopper; it isn’t a real customer’s words. Here is the audit of the listing at position 9. The store is withheld.
A. What we audited
| Field | Value |
|---|---|
| Product and variant | A 2.5-litre top-fill humidifier, store withheld. First returned variant, colour black |
| Route | Shopify Global Catalog search, top 10, position 9 |
| Request | “humidifier for dry winter air large room quiet top fill” |
| Buyer context | None set: no country, no signed-in buyer |
| Date and evidence link | Searched October 26, 2026; re-read from the saved response October 27. Private receipt kept internally |
B to D. Requirements, evidence and states
We used the four needs in the search, with the same room threshold as our coverage guide’s example brief.
| Requirement | Title and description say | Generated lists (Inferred) say | State |
|---|---|---|---|
| Humidifier | Title: “Ultrasonic Humidifiers for Bedroom” | Not needed | Supported |
| Large room: at least 450 sq ft | Nothing | “ideal for bedrooms and nurseries”; “fits easily in small spaces like bedrooms and nurseries” | Unknown. Words only in generated text, and possibly about footprint, not coverage |
| Quiet: a decibel figure | “Sleep peacefully with the low noise level” | “Noise Level: 26 dB”; “Ultra-quiet operation at 26 dB” | Words only in the description. A number only in generated text, with no setting or distance |
| Top fill | Title: “2.5L Top Fill” | “Easy top-fill design” | Supported, in the title |
The offer basics, read from the same record:
| Check | What the record says | Note |
|---|---|---|
| Price | USD 39.99 | Under the example budget in our coverage guide |
| Availability | Available | Sale-ready at the time |
| Ratings | Product 4.3 from 17 reviews; variant 4.6 from 21 | Two levels, two figures. Not a conflict to “fix”; they’re compiled differently |
| Policies | Refund and shipping policy links present | Content not audited here |
| Runtime | Description: “up to 25 hours”. Generated: “Up to 25 hours (low mode)” | The condition appears only in generated text |
E. The fix
Nothing on this record is wrong or doubled, so row 1 doesn’t apply. Row 2 does. “Quiet” is a stated requirement, the description answers it with an adjective, and a number already exists somewhere in Shopify’s data. On this shelf, four listings stated a decibel figure in their own description, and position 9 was one of two more that had one only in generated metadata.
| Card field | Entry |
|---|---|
| The change | Replace “low noise level” in the description with the measured figure, its setting and its distance, for example “26 dB on the lowest setting, measured at [distance]”, if the source says so |
| Owner | Whoever owns product copy for this SKU |
| Proof needed | The manual or a test report that states the figure and how it was measured. Shopify’s “26 dB” is a lead, not proof |
| Recheck | Re-read the catalog record after the edit and save it beside the original. The pages we read don’t say how quickly an edit reaches the catalog, so date the recheck and record what you find |
Queued behind it:
- Room coverage (row 3). Publish the rated coverage, with unit and conditions, from the manufacturer’s spec. If the unit is meant for small rooms, say so plainly. That turns an unknown into a true “no” for a 450 sq ft shopper, and that’s a return avoided, not a loss.
- Runtime condition (row 4). Move “on low” into the description’s own runtime line.
Notice what the audit didn’t recommend: longer copy, a new keyword, or more reviews. It found one sentence the shopper’s question depends on and the document that sentence needs.
What this audit is not
It isn’t a readiness grade. Shopify’s own listing quality indicator measures description completeness by word count, alongside image count, reviews, variant completeness and policy completeness (Managing agentic storefronts). That’s a useful hygiene check, but it answers a different question. A long description with no decibel figure can still count as complete. Our own first score failed the same way: about 65% of the points it awarded came from review depth, which was designed as only 25 of 100. It’s retired, and the audit of it explains why we don’t publish a replacement grade.
It isn’t a store audit. Vendor tools check different units. AgentReady describes auditing products, brand, merchant information, policies and structured data, then giving prioritized fixes. ARC lists the fields it says agents need and says page, feed and structured data should agree. Those are store-wide hygiene questions. This worksheet asks a narrower one: for this product and this shopper, which fact is missing?
It isn’t a pick prediction. Supported facts make a product checkable, not chosen. In our supplied-snippet simulation, vague copy was almost never picked: 3 of 180 Claude sessions, 0 of 60 on GPT-6.1-Sol, 6 of 60 on Gemini 3.8 Flash and 3 of 54 on Gemini 3.1 Pro, where random picking would be 25%. Facts got the product considered. It won when its facts answered a need the shopper stated that competing listings didn’t. That was a simulation, not live ChatGPT behaviour. OpenAI also says ChatGPT may generate simplified titles and descriptions, so put the deciding fact early in your own text.
It isn’t monitoring. One audit is one dated reading. Repeating it on a schedule is what our monitoring plan covers, and turning many audits into a team’s work queue is what the SKU dashboard guide covers.
How we checked
On October 27, 2026 we reopened the saved Shopify Global Catalog response for “humidifier for dry winter air large room quiet top fill”, taken October 26, 2026 with Shopify’s catalog command-line tool, top 10 results, no buyer context. For position 9 we read the title, the description with HTML removed, the generated metadata (unique selling points, top features, technical specifications), the price, availability, the product and first-variant ratings, and the seller’s policy links. Every quote on this page was checked against the field it is attributed to.
New observations from that record, beyond those already published:
- Position 9’s generated lists describe it as “ideal for bedrooms and nurseries” and say it “fits easily in small spaces like bedrooms and nurseries”. No field states a coverage area.
- The product rating was 4.3 from 17 reviews, and the first-variant rating 4.6 from 21.
- The seller record carried refund and shipping policy links.
We also re-read Shopify’s Global Catalog, agentic storefront and Catalog requirement pages, OpenAI’s shopping help page, and the AgentReady and ARC pages on October 27. Platform pages were checked on 2026-10-27; re-check them before acting. Six of the ten listings came from one seller, and we tested no product.
Your next step
Take your best-selling product and the question customers ask about it most. Fill in Part A, write three to five requirements in Part B, then read your received record into Part C, one quote per cell with its field named. Work down the fix order until you reach the first row that applies, and write that one card: the change, the owner, the proof and the recheck date.