An AI shopping readiness score can be useful for prioritizing work across a large catalog. It becomes misleading when it is presented as a prediction of whether ChatGPT, Gemini, or another AI shopping experience will recommend a product.
There is no single public, stable formula used by all AI shopping platforms to select products. OpenAI's public shopping explanation identifies factors such as relevance to the shopper's intent, product and merchant metadata, availability, price, quality, and whether a merchant is the maker or primary seller. It does not publish fixed weights or a universal score that merchants can optimize against. See Shopping with ChatGPT Search.
A score created by an audit tool is therefore a proxy measurement. It can measure technical conditions, information quality, consistency, or a sampled set of AI answers. It cannot reveal the platform's complete ranking system.
A more responsible audit separates four layers:
- Discoverability: Can the relevant channel access the product data or page?
- Comprehensibility: Can a system or shopper interpret the product's important facts and distinctions?
- Trust and entity clarity: Are the brand, seller, and product facts identifiable and consistent?
- Observed performance: What happened in a defined set of real AI shopping queries?
The first three are capability checks. The fourth is an empirical observation under a specific test design. Actual sessions, orders, and revenue are another measurement system, not something that a readiness score can replace.
Why one score can mislead
AI shopping is not one algorithm
Different channels may obtain product information through a merchant catalog, a product feed, web crawling, search indexes, or other providers. They also have different eligibility rules, shopping surfaces, and response behavior.
Shopify documents that supported AI channels can discover eligible products through Shopify Catalog and other sources, while each channel makes its own decisions about which products appear for a particular request. Shopify's product discovery documentation is a useful starting point, but it does not establish a universal optimization formula for all channels.
Static data is not a recommendation result
A product page can have complete structured data and still fail to appear in a given answer because the item does not fit the prompt, another product is a better match, availability or price differs, or the model uses different data. Conversely, an imperfectly marked-up product might still be found through a catalog or another route.
A high schema score does not prove a product will be recommended. A low schema score does not prove it will never appear.
The meaning of "readiness" varies by category
The information needed to choose a pair of running shoes differs from the information needed to choose a replacement industrial component, a skincare product, or a furniture item. Useful attributes, compatibility rules, regulatory details, and buying constraints differ by category.
An audit that gives every product the same universal checklist may score field presence while missing the information that actually separates relevant and irrelevant products for that category.
Scores age quickly
Catalogs change. Product feeds are updated. Models, shopping experiences, provider policies, and eligibility requirements change. A score from last month is a snapshot of a defined test, not a permanent property of the store.
A four-layer scoring model
If a score helps a merchant prioritize a large catalog, expose the components separately. Treat the following as an example audit framework that Miyeta can calibrate over time, not as an industry-standard formula or a known platform ranking model.
Score 1: Discoverability
Question: Can the selected channel receive the required product information through the route being tested?
Do not collapse product-feed distribution and public web crawling into the same check. For a Shopify-connected channel, Catalog eligibility and product-data mapping may be more relevant than a public crawler's ability to fetch a page. For public AI search, crawler access and page rendering may matter more.
An illustrative public-crawl score can use these operational checks:
| Check | Example weight | Evidence |
|---|---|---|
| Relevant crawler access or channel eligibility | 30% | Channel status, robots rules, verified request behavior where applicable |
| Product content accessible in the returned HTML | 25% | Rendered page capture and HTML inspection |
| Correct canonical and redirect behavior | 15% | Response chain, canonical tag, intended URL strategy |
| Sitemap coverage of canonical product URLs | 15% | Sitemap inspection against the product sample |
| Operational reliability and response performance | 15% | HTTP status, response time, failures over repeated tests |
These weights are a starting rubric for prioritizing technical checks. They are not known ranking weights used by an AI shopping platform. Page speed and response reliability are operational considerations; do not represent the suggested 15% weight as an observed factor in ChatGPT's shopping algorithm.
The crawler list must also be role-aware. OpenAI distinguishes OAI-SearchBot, GPTBot, and ChatGPT-User. Google-Extended is a robots.txt control token, not a separate HTTP User-Agent; Applebot-Extended is also a data-use control rather than a crawler that fetches webpages. See OpenAI's crawler documentation, Google's crawler reference, and Applebot documentation.
Gate logic: useful, but explicitly heuristic
Within a specific access route, some failures make downstream checks meaningless. If the relevant product cannot be accessed through that route, its content quality cannot be evaluated through that route.
For a transparent example, let D be the Discoverability score from 0 to 100. Define a gate factor:
Gate = min(D, 60) / 60
This sets the gate to 0 when access is completely blocked and caps it at 1 when the score reaches 60. The threshold of 60 is an illustrative operational choice, not an empirically proven boundary. A real audit should explain why it uses the chosen threshold and may change it after observing more data.
Do not apply this gate across unrelated routes. A blocked public crawler does not automatically mean a Shopify Catalog feed is blocked, so score the access route that matches the question being asked.
Score 2: Comprehensibility
Question: Does the available product data explain what the product is and when it fits a shopping need?
Inspect product information for the attributes that matter in the category:
- Product name, image, and factual description
- Price, currency, and availability where applicable
- Brand, SKU, GTIN, or MPN where relevant and available
- Materials, dimensions, capacity, ingredients, compatibility, or specifications
- Clear options and variant relationships
- Realistic use cases, constraints, and limitations
- Visible product details consistent with structured data
For Google Search, Product structured data can help make pages eligible for certain shopping and search presentations. Google recommends JSON-LD in general because it is easier to maintain, but also supports Microdata and RDFa. The required fields depend on the search feature being targeted. These are Google requirements and recommendations, not a universal list of AI shopping ranking requirements. See Google's Product structured data guide and structured data overview.
Do not score a product highly just because every possible field is filled in. A correct and useful set of attributes is better than a complete set of inaccurate, irrelevant, or unsupported claims.
Score 3: Trust and entity clarity
Question: Can a shopper or system distinguish the brand, merchant, and product, and can important claims be checked for consistency?
Review:
- Consistent brand and business identity
- Correct Organization or Brand structured data where applicable
- Accurate official profile links and
sameAsreferences, if used - Clear merchant/manufacturer relationship
- Consistent price, availability, product options, and policies across relevant pages and feeds
- Real reviews, specific product experiences, or substantiated use cases where available
- Product-specific descriptions rather than generic copy that obscures meaningful distinctions
Do not treat sameAs, a logo, or Organization schema as a known secret weight in an AI recommendation formula. They are identity and consistency aids. Nor should you claim that duplicated descriptions automatically cause a product to be demoted. There is no published universal rule establishing that outcome.
Instead, check whether the product information has enough factual detail to distinguish one product from alternatives. A generic description can be a quality problem because it does not explain product fit, even if a platform does not explicitly penalize duplication.
Score 4: Observed recommendation performance
Question: What happens when a defined set of shopping prompts is tested under recorded conditions?
This is the only layer that directly evaluates observed model outputs. It is still not a direct measurement of all shopper exposure or sales.
Build prompts around four types of intent:
| Test type | Example | What to record |
|---|---|---|
| Category request | "Recommend wireless headphones under $200 with multipoint Bluetooth." | Whether the product is included and its position |
| Comparison request | "Compare Brand A and Brand B for commuting." | Whether the product is described accurately and compared fairly |
| Problem or scenario | "Which lamp works on a small desk for video calls?" | Whether stated constraints match the product's actual specifications |
| Brand request | "What should I know about Brand X Model Y?" | Whether the correct entity and product are identified |
For each test, record:
- Included: brand or product appears in the answer
- First pick: appears first, if the answer has a ranked list
- Rank: position in the displayed order, where meaningful
- Citation: whether the answer links to or cites the relevant URL
- Competitors included: other products shown in the same result
- Stance: recommended, neutral, or discouraged, based on the answer's wording
- Test conditions: provider, model/version when visible, date, language, market, browsing mode, and exact prompt
Use multiple providers when possible and repeat a subset of prompts. A sample of 30 to 100 scenario prompts is a practical starting point for a manual audit, not a statistically validated rule. Small or repeated samples should be shown with an uncertainty range rather than presented as a definitive ranking.
A citation is evidence of source use in that answer. It is not proof of click-through, trust, or purchase. Inclusion rate and average position describe the test sample only, not every possible shopper query.
An example composite score
If one total is necessary for triage, keep the four underlying components visible. One illustrative formula is:
Composite Score = Gate × (0.25 × Comprehensibility
+ 0.25 × Trust
+ 0.50 × Observed Performance)
Where:
Gate = min(Discoverability, 60) / 60- Each component score is normalized from 0 to 100.
- Observed Performance has the largest weight because it uses measured outputs from the defined prompt sample rather than only structural checks.
This is a suggested audit heuristic, not a validated prediction model. The weights, threshold, prompt set, and normalization method must be documented. As more audits are completed, Miyeta should compare scores with subsequent observed results and revise the method if it does not help prioritize work.
For a small sample, report a range rather than false precision. For example, write "included in 7 of 30 sampled prompts across two providers" rather than claiming a universal 23.3% recommendation probability.
What a readiness score cannot measure
| Cannot reliably measure | Why |
|---|---|
| Whether a platform will recommend the product in a future answer | The complete recommendation system and its weights are not public, and results vary by context |
| The product's presence in every AI training dataset | Public output tests cannot reveal all training data or model memory |
| Real-time stock and price across every channel | Updates and data paths can have different timing; verify the source of each displayed value |
| Brand preference among human shoppers | A model-generated statement is not a measurement of consumer preference |
| Incremental orders or revenue from a product-information change | Requires commerce analytics and a credible comparison design |
| Performance in every country, language, or model | Tests must be segmented by market and provider |
| Competitors' private optimization strategies | Public answers reveal outputs, not another merchant's internal actions |
| Future performance after platform changes | Models, catalog sources, eligibility, and product ranking behavior can change |
How to present a responsible report
A useful audit should not end with "Your store scored 82/100, so AI will recommend you." Instead, it should state:
- What was tested: selected products, access paths, prompts, providers, market, dates, and settings.
- What passed or failed: concrete evidence such as missing attributes, mismapped fields, incorrect variants, or access errors.
- What was observed: product inclusion, citations, competitor appearances, and response variation in the sample.
- What should change first: a short list of controllable corrections with a clear validation method.
- What remains unknown: recommendation weights, unobserved exposures, causal sales effects, and changes outside the merchant's control.
This helps merchants separate technical prerequisites from information quality, observed model outputs, and actual commercial outcomes.
Final takeaway
A readiness score is useful when it organizes evidence and prioritizes action. It is harmful when it pretends to predict a black-box recommendation system with one precise number.
For Shopify brands, audit the actual distribution path, verify the product information, check the identity and consistency of the business, then test shopping prompts under repeatable conditions. Measure traffic and sales separately. Present the evidence and its limitations alongside every score.
That is a more trustworthy way to help a brand improve AI product discovery without promising a ranking or result that no independent audit can guarantee.