Build a price comparison you can actually trust

01Decide what comparable means

A pricing workflow should first establish that two records describe the same item and an eligible offer. Brand, model, size, pack count and condition can all affect comparability. A similar title is a candidate match, not a completed match.

Write rules for the comparisons the business actually makes. Does shipping belong in the comparison? Are member-only prices eligible? Are refurbished offers excluded? These choices determine the fields you need and should not be left to an analyst to infer from a single price column.

02Separate product identity from offer observations

Keep the product's stable attributes apart from the seller's offer and its observation time. One product can have several offers, and an offer can change without creating a new product. Preserve source-specific IDs and the mapping to your own catalogue.

Use explicit fields for regular price, current price, currency, shipping, tax treatment and eligibility. Missing shipping is not free shipping. A displayed discount percentage is not a substitute for the underlying values and the conditions under which the offer applies.

03Store snapshots before calculating changes

Keep the observation timestamp and collection context for each price. Compare the same eligible offer definition across snapshots. A change from one seller to another or from delivery to pickup can otherwise look like a price movement.

When an item disappears from a result set, mark it not observed in that run. Do not immediately conclude that it is discontinued or out of stock. Check whether discovery coverage changed, the page failed or the source returned a different regional catalogue.

04Quarantine surprising values

A sharp price change deserves inspection before it reaches an automated pricing rule. The cause may be a genuine promotion, a pack-size change, a parsing error or a currency mismatch. Keep the raw displayed value and source evidence so the reviewer can tell the difference.

Use thresholds appropriate to the category and business risk. A single global percentage rule can be too sensitive for volatile products and too lenient for stable ones. Validate the matching and extraction separately so a good parser does not hide a bad product match.

05Keep decision rules outside collection

A collection feed should report what was observed and how confidently it was matched. Repricing needs additional business rules such as margin floors, stock strategy and approval requirements. Begin with alerts or review queues and evaluate the quality of those signals before enabling consequential automatic changes.