Catalog Enrichment for AI: The Playbook

Catalog enrichment is the least exciting slide in any agentic-commerce pitch and the one that determines whether everything else works. Brand Agents, on-site AI search, and getting recommended by ChatGPT all read from the same place: your product data. If that data was written for humans skimming photos, the machines are working half-blind.

Why prose fails machines

A human reads "Crafted from buttery full-grain Italian leather that softens beautifully with age" and extracts: leather, full-grain, Italian. An agent reads the same sentence and extracts… a sentence. Unless material: full-grain leather exists as a structured field, the agent can't reliably answer "is this real leather?" — the single most common question for that product.

Structured attribute

A named, machine-readable field with a discrete value — e.g. material: cotton, sleeve_length: short, dishwasher_safe: true — as opposed to the same fact buried in a marketing paragraph.

The rule of thumb: if a shopper could ask about it, it must be a field. Prose is for delight; structure is for answers. You need both, but only one is machine-legible.

The three-part catalog audit

Before enriching anything, audit what you have. We run every catalog through three lenses:

1. Attribute coverage

For each category, list the attributes a shopper would ask about, then measure what percentage of SKUs actually have each one populated. Most mid-market catalogs come in between 40% and 60% coverage on the attributes that matter. The empty cells are your unanswerable questions.

2. Prose-trapped specs

Find the facts that exist only inside description paragraphs — dimensions mentioned in passing, materials named in marketing copy, care instructions hidden in a sentence. Every one of these needs to be extracted into a field. The fact is already there; it's just in a format the machine can't use.

3. Compatibility gaps

The most-overlooked and highest-value category: "works with", "fits", "suitable for", "pairs with". Agents lean on these relationships heavily because shoppers shop by them ("a case for an iPhone 16 Pro", "a filter that fits a Brita"). If your data has no compatibility relationships, the agent can't make the connection.

How to enrich, in priority order

You don't enrich a 2,000-SKU catalog uniformly — you'd run out of patience before you touched anything that matters. Sequence by revenue:

1. 1

   Hero SKUs first

   The 50–100 products that drive most revenue. Full enrichment: every attribute, every compatibility relationship, structured specs. This is where agent accuracy moves the P&L fastest.

2. 2

   High-traffic categories next

   The categories shoppers (and agents) hit most. Enrich to a consistent attribute schema across the whole category so comparisons work.

3. 3

   The long tail, to a baseline

   For the remaining SKUs, hit a minimum viable attribute set — enough that the agent never has to decline a basic question. Perfect is the enemy of shipped here.

4. 4

   Lock it with governance

   Set a required-attribute schema per category so new products can't enter the catalog half-empty. Enrichment without governance decays in a quarter.

One effort, three payoffs

Here's why catalog enrichment is the highest-leverage work in agentic commerce: the same structured data feeds three systems at once.

• Brand Agents answer accurately because the attributes exist — see Brand Agents explained.
• On-site AI search returns the right products because it can filter on real fields.
• Commerce GEO — off-site assistants like ChatGPT and Perplexity cite and recommend you because your data is machine-readable. See getting recommended by AI.

Common enrichment mistakes

• Enriching uniformly. Spreading effort evenly means hero SKUs get the same attention as dead stock. Sequence by revenue.
• Filling fields with junk. A populated-but-wrong attribute is worse than an empty one — the agent states it confidently. Accuracy over coverage.
• Skipping governance. Without a required-attribute schema, the catalog decays back to half-empty within a quarter.
• Ignoring compatibility. It's the hardest data to assemble and the highest-value for agents. Don't skip it because it's tedious.