A Claude citation to a store is not automatically a signal that the product information is correct or that the page will convert. In eCommerce, price, inventory, variants, reviews and return policy change faster than many editorial pages. The audit must first verify that the cited source describes the actual product and that the response preserves important commercial terms.
The goal is not to count domain occurrences and call them visibility. The goal is to identify where the citation helps, where it creates a false impression, and what finding can be fixed on the site, feed, or merchant data.
Start with real buying questions
Build your prompt set around product decisions: choosing a variant, compatibility, comparison, availability, delivery, returns and specifications. Very general brand questions can be useful on their own, but should not be mixed with those asking for transactional data.
For each prompt, note the intent and category. If the compatibility answers are good and the stock answers are poor, you have two different problems, not a single citation score.
Save the citation along with the supported claim
Don't just keep the URL. Capture the sentence or passage in which the source is used. Then check if the cited page actually supports that claim. A product page may be quoted correctly for specifications but incorrectly for availability if the data has changed.
In the audit, mark separately: relevant and correct citation, relevant but outdated citation, citation to an inappropriate page, third source better than own source, and answer without sufficient provenance.
Check price and stock at time of observation
In eCommerce, an answer can become outdated quickly. Save the state of the page or merchant data at the same time you collect the response. If the product was already out of stock, do not rate the quote as if it were current.
Google recommends structured Product and Offer data for information such as price and availability, and warns that dynamically generated information can be sensitive to crawling frequency. This doesn't prove anything about Claude, but it does show why one's own sources must be consistent and up-to-date.
Separate the product from the variant
A model may confuse a color, size or configuration with the parent product. Check that the quoted URL matches the described variant and that the identifiers are stable. In large catalogs, an almost identical name can hide differences in SKUs or GTINs.
If the page uses a single address for many variations, the audit should note how the selection is reflected in the content and structured data. It does not assume that the user or the system can automatically deduce the correct variant.
Control the relationship between feed and page
Merchant data, product page and internal system may temporarily end up in different states. A price that has been changed in the catalog but not propagated to the page creates a contradiction that an AI response can pick up from any available source.
Build sample products and check name, brand, identifiers, price, availability and URL. If these fields are not aligned, fixing the provenance is more urgent than rewriting the copy.
Don't treat reviews as specs
Reviews are useful for experiences, but may contain old versions, unusual usage or wrong information. If an answer uses a review to state a formal specification, the finding must be separated from an answer summarizing customer perceptions.
For the product, the specifications come from the controlled source. For experience, reviews can be community evidence. The two types of evidence should not be mixed.
Fake signals that look good
A citation to the homepage may look good in a dashboard, but may be poor if the question was about a specific product. A citation to a retired product may increase the count, but it makes for a bad experience. A category page quoted for exact price may hide the real variant.
Therefore, priority should not be calculated by frequency alone. It includes materiality: how many users can be misled and how quickly the information changes.
Audit decision tree
- Does the answer contain a verifiable claim about the product?
- Is there a citation or other visible provenance?
- Does the source match the product and variant?
- Was the price, stock and conditions current at the time of observation?
- If the source is third party, does the own website provide a clearer and more current version?
- Are there contradictions between feed, markup and visible content?
- Does Finding affect purchase information or just branding?
- After fixing, it repeats the exact prompt and keeps the result even if it doesn't change.
Remediation priorities
High priority: wrong product, wrong variant, out-of-stock price or stock, wrong safety claim or wrong compatibility. Medium priority: Citation to a less specific page when the information remains correct. Low priority: differences in wording that do not change the decision.
This pattern helps the team not consume entire sprints to cosmeticize appearances that have no impact.
Claim ledger
- FACT/EVIDENCE: Google documents Product, Offer, and merchant listing data for price, availability, and product identification in its experiences.
- FACT/EVIDENCE: Anthropic documents web search and citations in its tools, without guaranteeing that a particular merchant will be cited.
- PRACTITIONER GUIDANCE: citation classification and decision tree are audit methods proposed here.
- INFERENCE: consistency between feed, product page and markup can reduce source contradictions, but does not guarantee Claude citations.
- NOT PROVEN: that more citations directly produce more orders.
