Product comparison with preferences people can change

The starting point was collecting product information from retailer pages, storing structured price records and making them searchable for comparison. A product listing needed to become a record the application could work with, while retaining when it was observed.

PitchAI built a snapshot workflow that gives connector-provided data precedence and records changed or unchanged page observations. A content fingerprint helps distinguish a repeated page from a changed one. The resulting snapshots support comparison over time.

This distinction matters when a source page is refreshed. A new observation and a changed product page are different events; recording them separately gives the application more context than a single overwritten price field.

Later work added a recommendation workflow with explicit opt-in and visible preferences. Search and product-detail recommendations can use those saved signals and show reasons for their ordering.

A person can mark an item Useful or Not relevant, then undo the choice. An account view exposes active choices from the current consent decision, including the item, choice and time. Filters and pagination make that history navigable as it grows.

The feedback write checks the account's current consent decision and the recommendation that was shown. It verifies the owner and the consent decision attached to that recommendation before accepting a choice. A duplicate active choice is rejected.

Consent is checked again while holding a database lock for that user. This gives separate application workers a shared point at which to coordinate writes and resets. Feedback from an older consent decision is rejected, even if the user has since opted in again.

Reset and revoke behavior were exercised alongside delayed writes in database and browser tests. The checks covered multiple application workers, so the proof addressed more than one process remembering a local reset flag.

The recommendation and account-history work was implemented, reviewed and merged, with desktop and mobile browser evidence. These records establish a tested expansion of the comparison platform; they do not establish a production rollout of every recommendation feature or improved sales.

The project grew from collecting comparable product records into a workflow where people can inspect and revise the signals used to rank recommendations. The additional value is visible in the controls: reasons, active-choice history, Undo and reset.