Product category classification: how to assign products cleanly for retail media

Product category classification is the act of deciding where each product belongs in a category system. It sounds simple until a product could reasonably sit in three places, the source data is incomplete, or the same item is described differently by a supplier, a retailer and a media platform.

Retail media makes those disagreements expensive. Classification determines which products qualify for an audience, which sales count toward a campaign, which competitors appear in a benchmark and which items a planner can activate. If the assignments are wrong, the campaign can be executed exactly as booked and still answer the wrong commercial question.

This guide focuses on how a retailer or platform assigns products to clean groups and tests those assignments. The wider product taxonomy and SKU mapping layer defines the hierarchy, SKU identity and cross-system relationships that classification depends on. Once those product sets are reliable, they can support downstream uses such as category-affinity audiences.

Classification, taxonomy and SKU mapping are different jobs

A taxonomy is the category structure: for example, Grocery > Snacks > Bars > Protein Bars. Classification is the decision that a particular product belongs at one or more nodes in that structure. SKU mapping connects the retailer's product identifiers, variants and packs to the correct product and taxonomy records.

The distinction matters because each job fails differently. A taxonomy may be poorly designed even when every item follows it consistently. A sensible taxonomy may still contain thousands of misclassified products. Correct category assignments may also be attached to duplicate or obsolete SKU records.

GS1's Global Product Classification gives trading partners a common hierarchy for grouping products (GS1 Global Product Classification). A retailer may use GPC, a proprietary hierarchy or both. Google also maintains a predefined product category system for shopping feeds while allowing merchants to submit their own `product_type` hierarchy (Google product category guidance). These examples show the practical pattern: a controlled external classification can coexist with a business-specific internal view, but the relationship between them must be explicit.

Start with the decision the class must support

There is no universally perfect category assignment. The useful assignment depends on the decision.

A protein bar may be analyzed as a snack, sports-nutrition product, meal replacement or health-oriented food. The wrong response is to force every use case into one label. The better response is to establish a primary merchandise class, then add governed facets or mappings for alternate uses.

Before changing rules, state the jobs the classification must support:

• navigation and product discovery;
• assortment and category management;
• audience creation;
• competitive-set definition;
• campaign eligibility;
• sales and incrementality reporting;
• external feed requirements.

If one field is expected to perform all seven jobs, conflicts are inevitable. A hierarchy is good at rollups. Attributes and facets are better at describing characteristics such as vegan, premium, family pack, sensitive skin or compatible with a particular device.

Build a classification contract

A classification contract is a short, testable specification for each important category. It should include:

1. Scope: what the category means in plain language.
2. Inclusion rules: the attributes or product functions that qualify an item.
3. Exclusion rules: nearby products that must not enter.
4. Required evidence: which source fields support the decision.
5. Tie-break rule: what happens when several categories fit.
6. Allowed secondary mappings: where alternate analytical views are permitted.
7. Owner and review date: who resolves disputes and when the rule is reconsidered.

Consider a hypothetical "plant-based drinks" class. Inclusion may require a beverage whose primary base is oat, soy, almond, coconut or another non-dairy ingredient. Exclusions may cover dairy milk with plant additives, powdered supplements and yogurt drinks. The primary class could follow product form, while a separate dietary facet captures vegan status. That is much safer than classifying from title keywords alone.

Use evidence in a defined order

Product titles are useful but unreliable. "Kids," "light," "protein" or "natural" may describe positioning rather than product type. Classification should draw from the strongest available evidence in a fixed order, such as:

1. verified manufacturer or supplier type;
2. regulated or standardized product attributes;
3. ingredients, material or technical specifications;
4. package and usage data;
5. controlled brand-range mappings;
6. title and description text;
7. behavioral or co-purchase evidence for review, not silent reassignment.

Google's product-data specification requires accurate, correctly formatted product data and consistent price and availability information (Google product data specification). The same discipline applies internally. A classifier should not infer a product fact that contradicts the authoritative product record.

Machine learning can accelerate the work, but it does not remove the need for a contract. A model can propose labels and confidence scores. It cannot decide, without business rules, whether a dual-purpose item should be assigned by ingredients, shelf location, shopper use or reporting need.

A practical classification pipeline

A robust pipeline separates deterministic work from judgment.

1. Normalize the record

Standardize units, pack counts, brand names and controlled attribute values. Remove markup and obvious supplier noise without deleting meaningful distinctions. Preserve the source value and its provenance.

2. Resolve identity first

Confirm that the record represents the right product and variant. GS1 notes that GTINs uniquely identify trade items (GS1 GTIN standard). Classification performed before duplicate and variant resolution can create several conflicting labels for the same item.

3. Apply high-certainty rules

Use rules for cases with explicit evidence: a declared product type, a regulated class, a known brand range or an unambiguous attribute combination. Version the rules and record which one fired.

4. Score ambiguous records

Use a statistical or language model to rank candidate classes. Store the top candidates, confidence and features used. A confidence threshold should route uncertain or commercially sensitive products to review rather than forcing a label.

5. Review exceptions

Human reviewers need the product evidence, candidate categories and rule explanation on one screen. They should select a class, mark the reason and decide whether the case reveals a new rule or only a one-off exception.

6. Publish with lineage

The production record should carry the assigned class, classification method, source version, timestamp and rule or reviewer identifier. Without lineage, teams cannot explain why yesterday's audience differs from today's.

Test business impact, not only model accuracy

A global accuracy score can conceal the errors that matter most. High aggregate accuracy can still hide serious errors in a fast-growing category, promotional bundles or a brand's full range.

Use several checks:

• coverage: share of active products with a valid class;
• precision by important category: share of assigned items that truly belong;
• recall by important category: share of relevant items that were found;
• abstention rate: share routed to review;
• drift: change in assignments after feed, rule or model updates;
• commercial-weighted error: misclassification weighted by sales, campaign exposure or strategic importance;
• downstream reconciliation: differences in audience counts and reported sales caused by classification changes.

Retail media measurement guidelines emphasize defined product scope and consistent reporting boundaries (IAB/MRC Retail Media Measurement Guidelines). Classification QA is part of that boundary control. A campaign's promoted-product and halo-product sets should be frozen or versioned before launch, not reconstructed from a changed taxonomy after the result is known.

Worked example: classifying a protein snack range

Imagine a brand launches a range containing a single protein bar, a multipack, a powdered shake and a ready-to-drink shake.

A title-keyword classifier may put all four under "protein bars." A cleaner process separates form and purpose:

• the single bar and multipack share a bar product family but have distinct trade-item records;
• the powder belongs to powdered nutrition;
• the ready-to-drink item belongs to functional beverages or ready-to-drink nutrition under the retailer's primary hierarchy;
• all four may receive a governed "sports nutrition" analytical facet;
• only the bar items enter a campaign whose product scope is protein bars;
• a broader sports-nutrition audience may use the facet across all four.

This design avoids pretending that one category tree can express every commercial relationship. It also gives the planner a reproducible answer when asked which sales were included.

Governance questions buyers should ask

Before using classified product data for media, ask:

• Which hierarchy and version are active?
• What percentage of products were assigned by rules, models and humans?
• Which categories have the highest weighted error?
• Can a product have secondary analytical mappings?
• How are bundles, variants, private label and seasonal items handled?
• What happens when supplier data conflicts with retailer data?
• Are campaign product sets versioned at launch?
• Can the platform reproduce an old audience or report from the classification state used at the time?

These questions turn classification from a catalog-cleanup promise into an auditable operating process.

FAQ

Is product classification the same as product taxonomy?

No. Taxonomy defines the available categories and their relationships. Classification assigns products to those categories.

Should every product have only one category?

A product normally needs one governed primary class for rollups, but it may also need controlled secondary mappings or facets for valid analytical and activation uses.

Can AI classify an entire catalog automatically?

It can classify many clear cases and rank candidates for difficult ones. High-impact and low-confidence records still need rules, review and monitoring.

How often should classifications be refreshed?

Refresh when products, attributes, source feeds or category rules change. Monitor continuously for unclassified products and assignment drift rather than relying only on an annual cleanup.

What should retail media teams freeze before a campaign?

Freeze or version the promoted SKU set, category scope, halo set, exclusions and classification version used to build audiences and reports.

Bottom line

Clean product groups do not come from naming a hierarchy and hoping suppliers follow it. They come from explicit class definitions, evidence rules, identity resolution, exception handling and downstream tests. Treat classification as a versioned decision system. Then audiences and reports can explain exactly which products they included and why.

Ready to see how this works in practice?

Footprints AI helps brands and retailers measure what matters. See our customer success stories or get in touch to discuss your retail media strategy.