Book a live demo — see your catalog's review coverage →
Blog

GTIN, UPC, EAN, MPN, SKU: the identifier guide that makes review matching work

Reviews attach to products, so syndication is really a product identity problem. A field-by-field guide to the identifiers that matter and the ones that lie.

Matching
Data quality
Engineering
Aug 22, 2026 · 9 min read

By REVEX Engineering, Platform team

GTIN, UPC, EAN, MPN, SKU: the identifier guide that makes review matching work

Every review syndication failure we have investigated traces back to the same root cause: two companies disagreed about what product they were talking about.

This is a solved problem in theory and a mess in practice. Here is the working knowledge.

The identifiers, ranked by trustworthiness

GTIN (Global Trade Item Number)

The umbrella standard. UPC (12 digits, North America), EAN (13 digits, Europe), and ISBN are all GTINs of different lengths. A GTIN is issued through GS1 to a specific company prefix, and it identifies a specific sellable unit — one colorway, one size.

Trust level: high. When both sides have a GTIN and they match, you are looking at the same physical item. This is the only identifier that is globally unique by design.

Where it goes wrong: leading-zero truncation from spreadsheet exports (a 12-digit UPC becomes 11 digits and no longer matches), check-digit errors, and reused codes from companies that let their GS1 registration lapse. Always normalize to GTIN-14 with left-padded zeros before comparing, and validate the check digit.

MPN (Manufacturer Part Number)

The brand's own code for a product. Unique within a manufacturer, meaningless across them.

Trust level: high when paired with a verified brand. TR3-BK-105 matched to TR3-BK-105 is a strong signal — but only when both records agree on the manufacturer. MPNs collide across brands constantly.

Where it goes wrong: formatting drift. Distributors and retailers rewrite MPNs with different separators, casing, and padding. Normalize by uppercasing and stripping non-alphanumerics before comparing, and keep the original for display.

Style code / model number

A brand-level grouping above the variant: one style code covers all colorways and sizes.

Trust level: medium, and useful on purpose. Review content is usually about the model, not the specific size. A review of a size 10.5 black trail runner is genuinely informative for someone buying the 11 in grey — but rolling it up to a colorway page requires an explicit decision about how you present it, and honest disclosure of what the reviews are for.

SKU

Your internal code. It means nothing outside your four walls.

Trust level: zero for cross-company matching. SKUs are the most commonly populated and least useful field in any catalog export. Never match on SKU across organizations. Use it as a join key inside your own systems only.

Title and brand text

The fallback when nothing else exists.

Trust level: low, and dangerous because it feels confident. "Trail Runner 3" and "Trail Runner 3 Wide" differ by one token and are different products with different fit. "Pro Runner 2 (2026)" and "Pro Runner 2" may be the same product or a silent redesign.

A matching pipeline that does not embarrass you

Structure matching as tiers with different automation rules, not as one similarity score.

Tier 1 — Deterministic. Normalized GTIN equality, or verified brand plus normalized MPN equality. Auto-apply. Precision here should be effectively 100%; if it is not, your normalization is broken, not your logic.

Tier 2 — Strong heuristic. Brand match plus style code, or brand match plus a high-similarity normalized title with agreeing attributes (gender, width, model year). Auto-apply above a tuned threshold, and sample the output weekly to confirm the threshold still holds after catalog changes.

Tier 3 — Suggested. Everything else. Route to a human queue with both records shown side by side and the specific tokens that differ highlighted. Never auto-apply this tier, no matter how good the score looks.

The temptation is to widen Tier 2 to boost coverage. Resist it. Coverage bought with false matches shows up later as wrong sizing advice, higher returns, and a brand partner who revokes your grant.

Normalization rules worth writing down

  • Left-pad all GTINs to 14 digits and validate the check digit. Reject anything that fails rather than matching on a malformed code.
  • Uppercase and strip non-alphanumerics from MPNs and style codes; keep the raw value for display.
  • Strip marketing suffixes from titles ("New for 2026", "Free Shipping") before similarity scoring.
  • Treat width, gender, and model year as blocking attributes: if both records specify them and they disagree, the match fails regardless of text similarity.
  • Store the evidence for every match — which identifier fired, what score, which tier, and when — so a disputed match can be audited instead of argued about.

Practical advice for retailers

Populate barcodes on the variants that carry your revenue. In Shopify, that field already exists on every variant; treat it as required for anything you want covered by syndication. If your vendors do not supply GTINs in their price files, ask — most have them and simply do not include them in the retail feed.

Practical advice for brands

Publish your identifiers with your reviews. If you want your review content to travel to retail partners intact, the cheapest thing you can do is make sure every product record you expose carries GTIN, MPN, and style code. Matching quality at your partners is a direct function of the identifier quality you hand them, and every bad match is your brand being misrepresented on someone else's page.

Product identity is boring infrastructure. It is also the entire difference between review syndication that helps shoppers and review syndication that quietly lies to them.