How to Audit Product Evidence in AI Recommendations
A product recommendation shows an output, not the provider's complete decision process. You may see the prompt, answer, named products, attributes, and cited links. You usually cannot see every source considered, the internal candidate set, the ranking rules, or why one item was omitted. A defensible audit separates what was observed from what remains unknown.
Do not turn a plausible pipeline into a claimed fact
Retrieval, attribute extraction, identity matching, independent records, and answer synthesis are useful areas to inspect. They are not a universal five-stage mechanism proved for every provider. Features can use different indexes, catalogs, tools, models, account context, and policies, and those details can change. Treat each area as an evidence surface rather than a secret algorithm ClickRadius claims to know.
The answer is observable. The provider's complete candidate set and selection logic are not. Keep those evidence states separate.— ClickRadius Institute measurement rule
1. Define the observation
Preserve the exact question, provider, feature, available model identifier, account or locale context, date, complete answer, named products, exposed sources, and any error or missing response. One question answered once is one observation, not a stable recommendation rate. If the provider exposes no source for a statement, record the attribution as unresolved rather than guessing which page supplied it.
Use a frozen question set for comparisons and retain missing and failed samples in the denominator. A later answer change shows that the observed output changed; it does not, by itself, identify the cause.
2. Audit first-party product facts
Check the visible product name, model, price, currency, availability, dimensions, compatibility, identifiers, and policy details against the business's authoritative records. Then inspect any Product and Offer structured data for the same facts. Machine-readable markup can state publisher-controlled facts in a standard form where a consumer supports it. It does not establish that an AI provider crawled the page, accepted the markup, placed the item in a candidate set, or recommended it.
Unknown values should remain absent or explicitly unknown. Do not invent a price, availability state, rating, review count, identifier, or product relationship to make a graph look complete.
3. Check identifiers and record consistency
Compare exact names and applicable identifiers such as gtin, mpn, and sku across the first-party page and separately controlled retailer or catalog records. Consistency is a verifiable data-quality result. It does not prove that a particular provider performed entity resolution, merged those records, or assigned them any weight.
When two records disagree, preserve both sources and their observation times. Correct the record you are authorized to change, and leave the external record pending its publisher's action. Do not report the inconsistency as fixed until a fresh public readback confirms the changed surface.
4. Keep independent records independent
Retailer listings, review-platform pages, editorial comparisons, directories, and manufacturer pages are independently controlled evidence surfaces. Their existence, accessibility, and factual agreement can be observed. A mention is not automatically verified endorsement, a review aggregate is not automatically eligible markup for the represented business's own site, and the presence of an external page does not prove that an engine retrieved or trusted it.
For each source, retain who controls it, what it actually states, when it was read, whether the product identity is unambiguous, and whether the sampled answer cited it. That ledger supports auditing without turning correlation into a causal recommendation claim.
5. Compare the observed answer with the evidence
Record which products were named, which attributes were stated, which sources were exposed, and whether those statements match the retained source material. Classify an unsupported answer carefully: attributed-source conflict, conflicting sources, retrieval failure, model-generated text, provider change, non-attributable, or unresolved. Do not force every error into a single presumed source.
Repeat the frozen sample on its approved cadence. Separate the rate at which a product is named from factual accuracy, citation presence, and source agreement. These are different numerators and denominators.
An evidence checklist, not a recommendation scorecard
For any product you monitor, ask:
- Observation: Is the exact provider, prompt, date, response, and missingness retained?
- First-party facts: Do visible content, structured data, and authoritative business records agree?
- Identity: Are applicable product and variant identifiers consistent, with ambiguity recorded?
- Independent records: Is publisher control separated from factual agreement and exposed citation?
- Outcome: Are naming, accuracy, citations, and source agreement measured as separate results?
A complete checklist still does not establish eligibility or predict recommendation. It establishes which evidence was present, absent, conflicting, or unmeasured when the audit ran.
What ClickRadius does and does not measure
ClickRadius AI Readiness is a coverage-aware estimate across five evidence families: Off-site Authority, On-page GEO, Schema Engine, Entity / E-E-A-T, and Crawler Access. It does not mirror a shopping provider's private recommendation process. AI Presence is reported separately from Readiness and samples ChatGPT, Gemini, Perplexity, Claude, and Grok on the approved monthly cadence. Copilot is not currently supported by that automated matrix.
The score does not observe whether a shopping assistant retrieved or recommended a submitted product. External providers control crawling, indexing, retrieval, ranking, and recommendation. ClickRadius can preserve measured inputs and later answer observations; it cannot promise the external outcome or assign causal credit without suitable retained evidence.
Frequently asked questions
Do AI engines recommend products from a fixed catalog, or do they read the live web?
The answer varies by provider, feature, account, location, and date, and the complete source set may not be visible. Record the sources the response actually exposes and consult current provider documentation. Do not infer live retrieval, a frozen catalog, or a universal pipeline when that evidence is absent.
Why does a product with great specs still get left out of AI recommendations?
An omitted product does not reveal its cause. Check whether the product facts are accurate and visible, whether identifiers agree across controlled and independent records, and what the sampled response cited. Those checks can expose gaps, but they do not prove why the provider selected another answer.
Can I influence which products an AI engine recommends?
You can publish accurate, accessible product facts and correct inconsistent records you control. An external provider still controls crawling, retrieval, ranking, and recommendation. Measure later answers as dated observations; do not promise eligibility, selection, or a causal result.
Curious which readiness evidence is visible on the submitted product URL? Get your free AI Readiness Score for a five-family, 0–100 estimate, then see pricing when you are ready to review the measured gaps methodically. The score does not observe whether a shopping assistant retrieves or recommends the product.