Anatoliy Dankov
CEO
Quick answer:
When a SKU has no supplier description, text-based enrichment has nothing to search for. An image-based approach works instead:
Most product data enrichment guides assume a starting point: a manufacturer's PDF, a spec sheet, a competitor listing, or a review page. The AI finds that content, extracts it, and structures it into usable attributes. That works well when a product has any presence online.
A large share of retail catalogs don't. Private-label imports, small-batch toys, and unbranded apparel from overseas suppliers often arrive with a single line from the supplier and nothing else, something like "anti-stress toy, red, item 4471". A text-based enrichment tool searches for that product name and returns nothing, not because the AI is weak, but because there's nothing indexed to find.
This article explains how an image-based approach fills that gap, where else it's useful beyond missing descriptions, and how its cost compares to manual cataloging.
A large share of the retail assortment now comes from private-label and white-label sourcing, particularly in categories such as toys, accessories, and low-cost apparel. Store brands are not a niche: U.S. sales reached a record $282.8 billion in 2025, lifting their share to 21.3% of every dollar spent, according to Circana data released by the Private Label Manufacturers Association.
This creates a specific gap in the incoming product data. There's a name of sorts, a price, and an image, but no attributes, no description, no category mapping, and nothing a search engine has ever indexed. A supplier line like "anti-stress toy, red, item 4471" contains no material, no dimensions, no age range, no brand, and no usable detail for a product page.
For a retailer whose catalog is mostly branded goods, this is an edge case handled by a few people manually. For a retailer whose assortment is built on private-label sourcing, it's the default condition of nearly every incoming SKU. The volume makes manual cataloging the bottleneck, and it gets sharper during seasonal peaks. A new apparel collection or a holiday toy assortment can add thousands of undocumented SKUs within a few weeks, all of which need full product content before they can go live and arrive at once.
The cost of that gap shows up on both sides of the sale. A product published with thin or wrong data sells less, and it comes back more often. Returns are already expensive: the National Retail Federation estimates 19.3% of online sales were returned in 2025, and a product page that misleads on material, size, or audience feeds directly into that number.
Enriching a product with no existing description follows a different sequence than standard text-based enrichment. Rather than searching for a name and expanding what comes back, the process starts from whatever signals the incoming record already holds and builds outward from the product image. The steps below describe that sequence.
The starting point is the small set of data that does arrive with the product: the supplier line, the article number, the price, the assigned category, if there is one, and the product photo. Each of these is a signal, even when none of them is a full attribute. A price point narrows the likely product tier. A rough supplier category narrows the search space. The photo carries the most information of all, which is why it becomes the primary input when text signals run out.
With the product name returning nothing, the image becomes the query. The approach that works here checks the photo against a prioritized set of sources: a whitelist of trusted retailers, marketplaces, and manufacturer catalogs, paired with a blacklist that excludes low-quality or unreliable listings. Limiting the search to vetted sources matters more than search breadth. A wide, unfiltered search returns more matches but lower-quality ones, and a wrong attribute pulled from an unreliable listing costs more to catch later than it saves upfront.
Where the image matches a comparable product that does carry a full description, the relevant attributes are pulled and mapped to the record: category, color, material, style, and any structured detail the matched listing exposes. This is an extraction from a real source, not an invention. No attribute gets written that isn't traceable to a match, which keeps the output verifiable rather than speculative. Information that exists nowhere online can't be recovered by this or any method, since there's no source for any tool to draw on.
Every extracted attribute carries a confidence score and a source reference. A high-confidence match found across several independent sources is treated differently from a single low-confidence guess. The score is what makes the output auditable: a content manager can see why a product was classified as polyester rather than cotton, and which source that came from, instead of trusting an unexplained result.
Confidence thresholds decide what to publish automatically and what needs a human. A product enriched with high confidence across all key attributes can publish without review. One with a low-confidence match on a critical attribute routes to a content manager for a quick confirmation or correction. This keeps human attention on the cases that need it rather than on the whole catalog, which is what makes the process scale during a seasonal spike.
Image recognition earns its place in enrichment precisely when text runs out. Beyond filling in missing attributes, it classifies products along dimensions that the manufacturer information never captured. The subsections below cover the classification types that matter most for a product page and an ad feed.
The most immediate use is placing a product in the right category when the supplier gave none or gave a wrong one. Visual classification reads the product for what it actually is, a toy car versus a building set versus a plush toy, and assigns the category that the storefront's navigation and filters depend on. Correct categorization is the base layer everything else builds on, since a miscategorized product is invisible to shoppers browsing the right section.
The intended audience of a product often differs from what the incoming product records implies, and the image carries the signals that reveal it. The level of detail in the design, the finish, and the price tier together indicate whether a product targets young children, older children, or adult buyers. This matters for both on-site filtering and ad targeting, where the wrong audience assignment wastes spend on the wrong buyers.
Manufacturer gender tags reflect a decision made for the brand's home market and don't always match how a local shopper searches. Visual classification assigns a gender attribute based on the product's actual design cues: color palette, silhouette, print, and packaging. Storing this as a separate attribute from the manufacturer's own tag gives the storefront a locally accurate value to filter on without discarding the original data.
Themes and styles drive a large share of discovery in categories like toys and apparel, and they're rarely present in supplier data. Recognizing a licensed-style theme, a color story, or a design aesthetic from the image lets the storefront surface products through the theme-based searches and filters shoppers actually use. These attributes turn a generic listing into one that matches how people browse.
Matching a product to visually similar items already in the catalog serves two ends. It supports recommendations and "you may also like" placement, and it speeds enrichment itself by inheriting verified attributes from a close match that's already been cataloged. A new SKU that closely resembles an existing, fully enriched one starts from a strong baseline rather than from nothing.
Supplier attributes describe what a product is on paper. Marketing attributes describe how a customer looks for it. When the two diverge, the marketing attribute is what determines whether a product gets found and bought, which makes it the one worth getting right.
Toy safety regulations require an age label, commonly "3+", based on choking hazards and material testing rather than on who actually buys the product. A detailed die-cast collector's car priced well above the typical toy range is bought by adult collectors, not toddlers, whatever the safety label says.
Feeding the regulatory age into ad targeting or an on-site age filter misclassifies exactly the high-value products where accurate targeting matters most. The fix is to store two separate values, one for compliance and one for marketing, and to derive the marketing age from the image detail and price tier rather than from the safety label.
Large brands mark a wide range of products, unisex, a categorization made for the brand's home market. In markets where shoppers search "for girls" or "for boys" and skip a unisex filter by habit, an accurate unisex tag still reduces a product's visibility to the customers most likely to buy it. The label isn't wrong. It answers a different question than the one the storefront's search needs answered, which is why a locally derived gender attribute outperforms the manufacturer's own.
Suppliers organize products by how they manufacture and ship them. Shoppers organize products by how they think about buying them, and the two taxonomies rarely line up. A supplier category of "plastic figurines" maps to a dozen different things a shopper might search for, from a specific character theme to a gift for a particular age. Enrichment that classifies products around search behavior rather than supplier taxonomy is what closes the gap between what's in the catalog and what shoppers can find.
Google Shopping rewards complete, accurate product data, and penalizes gaps with lower relevance and wasted spend. Enriched attributes feed directly into feed quality and campaign structure. The subsections below cover where that connection is strongest.
Google Shopping supports up to five custom labels for segmenting products in a campaign. Enriched attributes, marketing age, audience, theme, style, price tier, give those labels something meaningful to carry. A feed built on enriched data can segment by buyer type or margin band rather than by whatever thin supplier category came in, which is what lets a campaign bid differently on products that deserve different treatment.
Not every product should be advertised. When a competitor's market price sits below a product's profitability threshold, showing an ad for it spends budget on a sale that loses money or never converts. Enriched data with reliable pricing and margin attributes makes it possible to identify these products and exclude them from the feed, rather than advertising the whole catalog indiscriminately and absorbing the loss on the products that shouldn't be there.
Feed visibility is most valuable on products where the retailer is actually competitive on price. Connecting price position to feed inclusion, so that products fall out of the feed when a competitor undercuts past a set threshold and return when the gap closes, concentrates ad spend where it converts. This turns the feed from a static export into something that reflects the current competitive position.
HootCore's PIM runs this enrichment sequence on catalogs with no supplier data, see how the AI Enrichment module works.
The steps described above work as a repeatable workflow rather than a set of manual decisions, and a PIM with a rule engine is where that workflow lives. A rule engine pairs a condition with an action and needs no developer to configure, which puts the workflow in the hands of the people who own the catalog.
The routing logic follows the enrichment sequence directly. If a text search returns no confident match, fall back to image-based enrichment. If image confidence falls below a set threshold, route the product to human review instead of publishing it. If a product's price exceeds a category threshold, flag its regulatory age label for a marketing-age override before it reaches an ad feed. Each rule is a condition and an action, set once and applied to every SKU that enters afterward.
Configured this way, the exception, deciding what to do with a product nobody described, becomes a standard branch every new SKU passes through automatically. The rules apply identically whether ten products enter in a week or ten thousand enter over a weekend during a seasonal peak, which is the condition under which a manual process breaks down. The practical starting point is auditing a sample of the current catalog for how many SKUs would fail a text search entirely. That number sets the scale of the problem before any tool or rule structure gets chosen.
Most enrichment tools can attempt a catalog. The useful questions are about what happens at the edges, where a private-label assortment actually lives. Before committing to a tool, these are worth asking directly.
Many tools enrich by expanding existing text and stall when there's nothing to expand. A tool built for private-label catalogs has to start from the image, not the name.
Image-based matching is the difference between enriching a documented product and enriching one that exists nowhere online. Without it, undocumented SKUs still fall to manual work.
If every routing rule needs a developer, the workflow won't keep pace with the catalog. The people who own the products should be able to set conditions and actions themselves.
A retailer selling across markets needs attributes and descriptions in each market's language, generated from the same enrichment pass rather than as a separate translation project.
Enrichment that doesn't feed cleanly into the systems already running the catalog and the orders creates a parallel data source to reconcile. Integration with the existing stack is what keeps enriched data usable end-to-end.
When a product arrives with nothing but a photo and an article number, expanding an existing description isn't an option, because there's no description to expand. What works instead is reading the product itself, from its image, its price, and its place among comparable goods, and building the data outward from there.
That capability matters most for the retailers whose catalogs are hardest to enrich: private-label sellers, importers of unbranded goods, and anyone whose suppliers stop at a photo and a number. For them, enrichment that depends on existing online content solves the easy half of the catalog and leaves the hard half untouched. Enrichment that starts from the product itself is what closes the gap.
See it in your own catalog. Book a demo and bring a batch of your hardest SKUs, the ones with nothing but a photo and an article number.
Talk to our team and see how HootCore fits into your existing stack, from product data management to order fulfillment.