Canonical product identity: one product, not forty listings
Every price question starts with an identity question. If two listings are not the same product in the same configuration, comparing their prices produces a number that looks precise and means nothing.
- 4
- 0.55
- 6
identifier, model+config, fuzzy, unresolved
below this a listing never joins a comp set
model, storage, color, region, kit, generation
The problem: titles are marketing copy, not identifiers
A single camera body appears on marketplaces as "Sony A7IV Body Only MINT", "a7 iv 33mp full frame ILCE-7M4", and "Sony Alpha 7 IV + 2 batteries + 128GB card". These are three different economic objects — one is a bare body, one is a bundle — but a naive text search treats them as interchangeable and reports a price spread that is actually a bundling difference.
Marketplace category trees do not solve this. They are optimized for browsing, not for equivalence. A category can contain body-only listings, kits, and broken units side by side.
How resolution works
MarginMap resolves each raw listing to a canonical product variant through an ordered cascade. Higher tiers are trusted absolutely; lower tiers carry a confidence penalty that flows into every downstream score.
- Tier 1 — structured identifier
- GTIN, UPC, EAN, MPN or platform catalog ID present and valid. Deterministic match, confidence 1.0.
- Tier 2 — model plus configuration
- Extracted model code plus resolved variant axes (storage, region, generation). Confidence 0.85–0.95.
- Tier 3 — fuzzy title match
- Normalized token overlap against known variant aliases, penalized for unmatched tokens that indicate bundles or accessories. Confidence 0.55–0.85.
- Tier 4 — unresolved
- Listing is retained as raw evidence but is excluded from comps, medians and scores. It never silently inflates or deflates a market estimate.
Bundles are separated, not averaged
Extra batteries, memory cards, extended warranties and shipping insurance all move the observed price without changing the product. MarginMap detects accessory tokens and either splits the bundle into a base variant with an itemized add-on set, or drops the listing to Tier 4 when the split cannot be made honestly.
This matters most on the sell side. A reseller who prices against bundle comps consistently overestimates what a bare unit will fetch.
Why confidence travels with the number
Identity confidence is not discarded after matching. It is one of the inputs to the data confidence factor in both the buyer score and the reseller evaluation, and it is displayed in the evidence drawer next to every offer. A median built from Tier 3 matches is labelled as such rather than presented with the same authority as one built from verified identifiers.
Frequently asked
Keep reading
Run one product through it
Open the workspace, search in plain language, and open the evidence drawer on the first number you doubt.