A marketplace operator needs to match seller listings across two product catalogs. Both catalogs have product names and product photos, but the text descriptions are wildly inconsistent. Catalog A calls it “Casual Leather Belt Men’s Brown.” Catalog B calls it “Genuine Full Grain Leather Dress Belt — Cognac, 34.” They’re the same belt. The photos prove it — same buckle, same stitching, same leather grain pattern.

But a text-only matching algorithm sees “Casual” vs “Dress,” “Brown” vs “Cognac,” and the presence of a size in one but not the other. It scores this pair as a low-confidence match at best, a non-match at worst.

The product images contain the attributes that would make this match obvious: the buckle style, the leather type, the color (which is cognac, not brown or generic “leather”), the stitching pattern, the brand stamp on the back. These attributes aren’t in the spreadsheet. They’re in the photos.

Why text-based product matching falls short

Product descriptions are written by different people with different vocabularies, different priorities, and different levels of detail. The same physical product can be described in ways that share almost no words.

Two listings for the same product — text fails, images match
Field Catalog A Catalog B
Product name Casual Leather Belt Men's Brown Genuine Full Grain Leather Dress Belt — Cognac, 34
Category Accessories Men's > Belts > Leather
Price $39.99 $42.00
Brand — —
Material — Genuine leather
Color Brown Cognac
Size — 34
Image belt_photo_A.jpg belt_photo_B.jpg

Text descriptions overlap minimally. Both images show the same belt — same buckle, leather grain, and stitching pattern.

The text fields have almost nothing in common. “Casual” and “Dress” are opposites. “Brown” and “Cognac” are technically different colors. The category taxonomies don’t match. One lists a size, the other doesn’t. Neither includes a brand.

A human looking at both photos would match them instantly. The challenge is getting an automated system to extract what the human sees and turn it into structured data the matching algorithm can use.

What images reveal that text doesn’t

A product image is information-dense. A single photo can contain dozens of attributes that the text listing never mentions.

Brand identifiers. Logos on the product, brand stamps on leather goods, label designs on clothing, wordmarks on electronics. Even when the text listing omits the brand, the product photo often shows it clearly.

Exact color. Text descriptions use approximate color names. “Brown” could be chocolate, espresso, tan, cognac, or chestnut. “Blue” could be navy, cobalt, royal, powder, or slate. The photo shows the precise shade, and AI can describe it accurately — “warm cognac brown with reddish undertones” disambiguates in a way that “brown” never can.

Material and texture. Visual texture is immediately apparent to both humans and AI. Full-grain leather looks different from bonded leather. Brushed metal looks different from polished. Cotton jersey looks different from polyester interlock. These distinctions matter for matching — a cotton crew neck and a polyester crew neck are different products even if the text description is identical.

Design details. Button count on a shirt. Stitching pattern on a glove. Buckle shape on a belt. Tread pattern on a shoe. These micro-attributes are almost never in text listings but are clearly visible in photos and highly distinctive.

Physical condition. New with tags, new without tags, open box, refurbished, used. The product photo reveals condition details that sellers may describe inconsistently or omit entirely.

Packaging and accessories. What’s included in the box — cables, adapters, manuals, cases. A photo of the product in its packaging reveals the complete offering, while the text listing might say “with accessories” or nothing at all.

Multi-field extraction from a single image

The power of AI attribute extraction is that a single image can yield many structured fields in one pass. You don’t need separate models for brand detection, color identification, and material classification. One multimodal AI call reads the image and produces all attributes simultaneously.

Attributes extracted from a single product image
Attribute From text listing From image extraction
Brand — Timberland
Product type Men's Boots 6-inch premium waterproof boot
Color Wheat Wheat nubuck (golden tan)
Material (upper) — Premium nubuck leather
Material (sole) — Rubber lug outsole
Closure — Lace-up, 7 eyelets
Collar — Padded leather collar
Waterproofing — Seam-sealed construction visible
Condition New New, with original hang tag
Logo placement — Embossed tree logo on upper shaft
Sole pattern — Lug tread for traction

The text listing provided 3 attributes. Image extraction added 8 more.

The text listing had three useful fields: product type, color, and condition. The image extraction added eight more: brand, specific product type, exact material for upper and sole, closure type, collar detail, waterproofing evidence, and logo placement.

With 11 attributes instead of 3, this product can be reliably distinguished from similar boots and confidently matched against the same boot in another catalog — even if the other catalog describes it completely differently.

Extraction prompt design

The quality of extracted attributes depends heavily on how you ask for them. Vague prompts produce vague results. Specific prompts produce specific, matchable data.

Vague prompt: “Describe this product image.”

This returns a narrative description: “A pair of wheat-colored boots with brown laces and a rugged sole.” Useful for embeddings, but not structured enough for precise matching.

Specific prompt: “Extract the following attributes from this product image: brand (from any visible logo or label), exact color (specific shade, not generic), material (upper and sole separately), closure type, and condition (new/used/refurbished with evidence).”

This returns structured fields that can be directly compared against another product’s attributes. Brand can be string-matched. Color can be embedded. Material can be compared. Condition can be filtered.

Multi-product awareness: “This image shows a single product. If multiple items are visible (e.g., a pair of shoes with an included bag), describe only the primary product. List accessories separately.”

Edge case handling prevents the AI from conflating the product with its packaging, display stand, or included accessories.

Extraction quality and edge cases

AI attribute extraction is highly accurate for well-photographed products, but accuracy varies by attribute type.

Extraction confidence by attribute type
Brand (when logo visible) Very high — logos are distinctive
94%
Primary color High — visual color is unambiguous
92%
Product category High — product type is visually clear
91%
Material type Medium — some materials look similar
78%
Exact dimensions Low — no reference object in most photos
42%
Weight Very low — not visually determinable
18%

Confidence represents the percentage of cases where the extracted value is accurate enough for matching. Dimensions and weight require reference objects or prior knowledge.

The pattern is clear: attributes that are visually distinctive — brand logos, colors, product types — extract with high confidence. Attributes that require inference — material identification from texture — are moderately reliable. Attributes that are not visually determinable — weight, exact dimensions — are low-confidence and should be sourced from text fields or spec sheets rather than images.

The practical implication: use image extraction for the attributes images are good at (brand, color, design, condition) and text extraction for the attributes text is good at (model numbers, dimensions, specifications). The combination covers more ground than either alone.

Common edge cases

Blurry or low-resolution images. AI extraction degrades gracefully — it returns fewer attributes with lower confidence rather than hallucinating details. A blurry logo might be identified as “brand text present but unreadable” rather than guessing.

Multiple products in one image. Catalog photos sometimes show product bundles or lifestyle shots with multiple items. Prompt design helps: specify “extract attributes for the primary product only” and the model typically focuses on the dominant item.

Stock photos vs. actual product photos. Generic stock photos (e.g., a generic picture of “headphones”) provide less extractable information than actual product photography. The AI can often detect this — “This appears to be a generic/stock product image rather than a specific product photo.”

From extracted attributes to matching

Once both datasets have been enriched with image-extracted attributes, matching uses the same techniques as text-based matching — but with more signals to work with.

String matching on brand. Brand extracted from one catalog’s images can be exact-matched or fuzzy-matched against brand from the other catalog. “Timberland” matches “Timberland” with certainty.

Embedding matching on descriptions. The full set of extracted attributes, concatenated into a natural-language description, produces a rich embedding. “Timberland 6-inch premium waterproof boot, wheat nubuck leather, rubber lug sole, lace-up” embeds near the same product described as “Timberland Premium Boot 6 inch Wheat” — even though the word overlap is partial.

LLM confirmation with side-by-side images. For borderline matches where the text and embeddings aren’t conclusive, the final step is showing both product images to the AI side by side: “Are these photos of the same product?” This visual comparison catches cases that attribute-level matching might miss — same product from different angles, same product in different lighting, same product with and without packaging.

The pipeline in Match Data Studio

The end-to-end workflow:

  1. Upload two CSVs — each with a column referencing product image filenames.
  2. Upload product images for both datasets.
  3. Mark image columns as “file” type.
  4. Create AI enrichment rules — “Extract brand, color, material, product type, condition, and notable design features from this product image.”
  5. Configure embeddings on both the text columns and the file columns. Image descriptions are auto-generated and embedded.
  6. Set similarity thresholds — how similar do embeddings need to be to be considered a candidate match?
  7. Add LLM confirmation — include the image columns so the AI can compare photos side by side for borderline pairs.
  8. Run the pipeline and download the matched results.

The extracted attributes become permanent columns in your enriched dataset. You can download them alongside the match results — getting both the matches and the structured product data as output.

For the complete walkthrough of combining image attributes with text matching in a single pipeline, see our guide on matching with images and attributes.

Your product images are your richest data source. Let AI read them.

Start extracting attributes →

Keep reading