An effective review queue turns every doubtful match into a traceable decision. Place in it the supplier–marketplace pair, the identifiers, titles, images, brand, variant, packaging, and evidence used. Show matching elements and contradictions separately. The reviewer must be able to confirm, reject, request evidence, or declare the information insufficient, with a standardized reason and a comment. Prioritize by potential impact and uncertainty, without an opaque score. A decision must record the author, the time, the rule version, and the observed data. Feedback into matching serves to correct future rules; it must never silently rewrite a file that has already been reviewed.
A false positive is a decision, not just a bad score
Two listings can share words and an image while referring to different products: capacity, color, unit count, region, condition, or generation can vary. Conversely, very different titles can describe the same item. A review queue exists to focus human attention where automatic evidence is incomplete or contradictory.
GS1 defines the GTIN as a key used to identify a trade item. A valid GTIN assigned at the right level is therefore an important piece of structured evidence. However, you must still verify that the identifier truly belongs to the variant and packaging being compared. The check-digit control catches certain data-entry errors; it alone does not prove that the code was assigned to the displayed product.
When the EAN or GTIN is missing, matching by image and name only produces a bundle of clues. The image may be generic, cropped, or shared across several variants. The title may omit a decisive characteristic. The queue must make these limits visible.
The minimum file for a match
| Element | Expected content | Why keep it |
|---|---|---|
| candidate identifiers | source, supplier reference, marketplace, ASIN or identifier | retrieve both listings |
| observed GTIN/EAN | value, location, and validation result | distinguish presence, format, and assignment |
| raw and normalized titles | versions before and after transformation | explain textual matches |
| images and provenance | URL or fingerprint, view type, date | review the visual evidence |
| attributes | brand, model, color, size, capacity, lot, condition | detect contradictions |
| rule and version | criteria applied when the candidate was proposed | reproduce the result |
| supporting elements | explicit evidence of a match | avoid an unexplained score |
| blocking elements | conflicts or missing data | guide the review |
| decision | confirm, reject, escalate, insufficient | close the case clearly |
| author and timestamp | reviewer, date, comment | keep accountability |
The W3C's PROV-O model provides provenance concepts to link entities, activities, and agents. Without requiring its technical adoption, this separation offers good discipline: the compared data are entities, matching and review are activities, and the system and the reviewer are identifiable agents.
Reproducible review procedure
1. Freeze the observed file. Keep the data and rule version that created the candidate. 2. Verify commercial identity. Compare GTIN, manufacturer part number, brand, and model before visual similarities. 3. Compare the variant. Check size, color, capacity, region, compatibility, and condition. 4. Compare the packaging. A single unit, a lot, or a multipack are not interchangeable. 5. Examine the raw text. Go back to the original titles to spot information removed by normalization. 6. Examine the images. Look for visible characteristics, while noting that a generic image is not decisive proof. 7. Classify contradictions. Use stable reasons, e.g. different variant, different lot, or insufficient identity. 8. Make an explicit decision. Confirm, reject, escalate, or mark the file insufficient. 9. Document the justification. Link the reason to the evidence, with author and timestamp. 10. Feed back without rewriting. Propose a rule improvement for future cases while keeping the original result.
Mapping a product across Amazon marketplaces illustrates another source of ambiguity: an ASIN is interpreted within a marketplace's context. The queue must therefore keep the relevant market instead of assuming a global identity.
Prioritizing without hiding uncertainty
Priority can combine the impact of a wrong match with the difficulty of the decision, but each component must stay visible. A candidate tied to a large quantity may be reviewed before a candidate with no immediate stakes. A capacity contradiction can also justify a quick review. No universal coefficient can be recommended without context.
A simple method is to define explicit classes: contradictory structured evidence, missing critical attribute, ambiguous image, or routine confirmation. Within a class, age can break ties between files. This approach is easier to audit than a single number whose inputs no one knows.
The queue must also avoid clogging up. Group strictly identical candidates, close duplicates with a reference to the original decision, and measure the backlog by state. A corrected rule can requalify unreviewed cases, but must not retroactively change human decisions.
Hypothetical decision example
Assume two fictional listings sharing the same model name and a similar image. The supplier listing states a lot of two units for 30.00; the marketplace listing describes one unit for 24.00. The amounts and products are invented solely to illustrate the check.
Even though the brand and model match, the packaging contradicts the offer's identity. The reviewer chooses "rejected — different quantity," keeps both texts, images, identifiers, and the rule used. No margin is inferred, since the compared units are not the same.
If the marketplace listing did not state the number of units, the appropriate decision might be "insufficient information" rather than "confirmed." Additional evidence, for example structured data from the manufacturer or the marketplace, would then be requested.

This screenshot is a redacted public view. It contains no account data and does not demonstrate the existence of the proposed authenticated feature.
What ArbitragePro+ could automate / what the seller must still verify
The specification:file-revue-correspondances feature is a proposal, not a confirmed feature of the application.
| Automation proposed for ArbitragePro+ | Verification that remains the seller's |
|---|---|
| gather candidates and their evidence | verify identity, variant, and packaging |
| flag matching and contradictory attributes | interpret text or image ambiguities |
| deduplicate and classify files | set priorities based on business risk |
| log decision, reason, and rule version | own the decision and request missing evidence |
| propose corrections for future matches | validate any rule change before applying it |
The proposed automation would prepare the file and preserve the decision. It would not replace product verification or authorization to sell.
Reviewer checklist
- Both listings and their markets are identified.
- The GTIN is checked at the right packaging level.
- Brand, model, and manufacturer part number are compared.
- Variant, condition, and compatibility are checked.
- Quantity and sales unit are identical.
- The raw titles remain visible.
- The images' provenance is known.
- Contradictions are separated from missing data.
- The decision has a reason, an author, and a date.
- The opportunity export keeps the review result.
Official sources
Sources consulted on August 17, 2026:
- GS1 — Global Trade Item Number
- GS1 — check-digit calculation and validation
- GS1 — Product Image Specification Standard
- W3C — PROV-O, official provenance ontology
