Product Data Audit for Ecommerce Search: A Field-by-Field Procedure

A random sample alone can miss the records with the highest search risk. A list of top sellers alone can describe only the records your team already maintains well. Combine demand, known failures, recent change, structural complexity, and long-tail coverage.

Keep stable product and variant IDs in the worksheet. Titles and SKUs can change during the repair, so they are not reliable row keys.

Identity fields answer “which exact item?” Titles answer “what is this?” Do not use one as a substitute for the other. An internal code can be a strong exact-search key while remaining a poor result-card title.

Look for blank identifiers, duplicated values, IDs attached to more than one variant, generic titles, title collisions within a product family, and formatting changes that create a second record instead of updating the first.

A parent product can be valid while one variant is missing a SKU, option, image, price, stock, or compatibility value. Sample at the variant level and keep the expected variant ID in every query test.

For Shopify stores, the variants documentation checked July 28, 2026 describes each option combination as a variant. Your search implementation may index products, variants, or a mixed projection, so verify which identity the response returns and which identity the card preserves.

Classification failures create both missing and irrelevant browse paths. Record the product’s standard category, custom product type, manual collections, automated collection rules, and compatibility families. These fields are related, but they are not interchangeable.

Shopify’s current taxonomy documentation distinguishes its standard product category from the store’s custom product type and links category assignments to category metafields. Audit the actual rule that placed each product. A correct label with a broken automated rule will drift again after the cleanup.

Completeness needs a denominator. Measure a field only across products where the category contract says it applies. Otherwise a valid not-applicable state looks like a defect.

Keep unknown, invalid, and conflicting values separate. They may all render as blank, but the repair owner and risk differ. An unknown care instruction needs enrichment. Conflicting voltage needs source resolution before the product should satisfy a constrained query.

Case, punctuation, whitespace, and supplier abbreviations can fragment one logical value. Different shades, fitment families, units, and sub-brands can look similar while remaining distinct buyer choices.

Create a proposed mapping table, have the category owner approve it, then test both a query and the filter generated from the source field. Keep canonical values separate from display labels and query aliases.

Price, stock, publication, product copy, and variant changes can have different pipelines and different acceptable delays. Choose non-critical fixtures for each event type, make one change, and record the first correct state at every layer.

A single storefront timestamp is not enough to locate the lag. Capture the source update, event receipt, transformed record, successful index write, search response, and rendered result.

Choose a rare controlled value that appears on one known product or variant. Trace the value from source to rendered result. Repeat the test for the predictive dropdown, full results page, filtered browse path, or API surface your shoppers actually use.

On Shopify, the current regular storefront-search documentation and predictive-search documentation list different searchable product properties and request behavior. A SKU that works after pressing Enter can still fail in the dropdown without proving that the variant record is empty.

Do not label every blank field critical. Prioritize the confirmed failure, how often the buyer job occurs, how much of the applicable catalog it reaches, and whether one source or validation change prevents recurrence.

Keep confidence separate from impact. A severe-looking symptom with an unproven cause should receive an investigation owner, not a mass catalog rewrite.

If the source record or field contract is wrong, use the catalog-quality framework to repair the model and prevent recurrence. If the source is correct but the search document is stale or shallow, use the indexing health procedure.

If the expected product is retrieved but ordered poorly, use the relevance evaluation system. If the filter misses valid products, use the faceted-navigation guide to inspect filter applicability, counts, and result state.