Picky.expertPicky.expert

Methodology

A technical look at how raw seller data becomes the structured rankings on Picky.expert. If you're new here, start with How We Rank for the plain-language version.

1. Ingestion

We continuously pull product listings and offers from retailer sources. Each raw row is stored verbatim with a timestamp and the source URL, so any downstream claim can be traced back.

2. Normalization & deduplication

The same product is often listed under dozens of titles ("Apple iPhone 15 128GB Black Unlocked", "iPhone 15 — 128 GB"). We normalize names with an AI model constrained to the format Brand + Model, then auto-merge entries with matching normalized names as variants. A canonical record is created; the other listings become variants attached to it.

3. Attribute extraction

Specs are extracted from listing text and manufacturer pages by structured prompts that return JSON conforming to a per-category schema. Extracted values are then validated against allowed types and ranges; out-of-bounds values are rejected, not silently coerced.

4. Taxonomy maintenance

Product categories and tags are kept consolidated. Synonyms (e.g., "mobile phone", "smartphone") collapse to a single canonical value with the rest stored as aliases that 301-redirect. Action-style tags ("comparison", "troubleshooting") are blocked at extraction time so the taxonomy stays focused on product entities.

5. Refresh cadence

  • Offer prices: refreshed on a rolling cycle, typically daily.
  • Product attributes: refreshed when source listings change.
  • Ratings: re-pulled on the same cycle as offers.
  • Taxonomy: maintenance pass runs continuously in the background.

6. Human review checkpoints

AI is fast but not infallible. Human reviewers spot-check normalized names, validate new categories before they become landing pages, and review every community-submitted edit before it merges into the canonical record.

7. Known limitations

  • Coverage depth varies by category — categories with fewer listed products produce fewer comparison pages.
  • Prices reflect the last refresh, not the live retailer price at click time. Always confirm on the seller's site before purchasing.
  • Ratings inherit any bias present in the retailer source they came from.

Find an error in the methodology or its output? Tell us on the contact page.