How We Enrich 10 Sample SKUs Before You Commit to Catalog Work

A free 10-SKU enrichment sample answers one operator question: can we turn the facts already sitting in your product copy, manuals, and support answers into structured fields that Brand Agents, on-site AI search, Rufus, and off-site assistants can actually use — without committing the whole catalog first? It is not a scored PDF that says “your data needs work.” It is a worked before/after on ten real products: which attributes were missing, which values came from prose versus a supplier sheet, and what a publishable metafield or Amazon attribute row looks like for your category.

This is the GigaCommerce method behind the Catalog Enrichment for AI sample offer. It sits next to the broader catalog enrichment playbook as the proof step — show the work on ten SKUs, then decide scope. Operators in the $100K–$10M band use it when leadership wants evidence before budget, when a Plus upgrade for Brand Agents is on the table, or when Amazon listing quality keeps bouncing on incomplete product-type attributes.

Why a 10-SKU sample exists

Most mid-market catalogs fail AI shopping the same way: the shopper-facing story is fine, and the machine-facing contract is empty. Dimensions live in a paragraph. Material is an adjective. “Works with” is buried in a Q&A reply. Compatibility is a PDF nobody indexed. A Brand Agent that only has title, price, and a marketing blurb will hedge, invent, or escalate — and every wrong confident answer becomes a return or a bad review.

Full enrichment across hundreds or thousands of SKUs is real work. Teams that skip a sample either under-scope (they discover the category needs twenty required fields after kickoff) or over-scope (they buy a PIM before they know whether Shopify metafields would have been enough). The sample compresses that discovery into one week of evidence.

It also separates two different problems. Problem A: the facts exist somewhere and just need extraction and typing. Problem B: the facts never existed — no supplier sheet, no measured dimension, no verified model list. Problem A is enrichment. Problem B is product ops. A 10-SKU pass tells you which problem you actually have on the SKUs that matter.

How to pick the ten SKUs (selection is the method)

Bad samples produce false comfort. Ten bestselling, simple, single-variant SKUs always look “mostly complete.” Agents break on the awkward products. Use a deliberate mix.

Revenue heroes (4–5 SKUs)

Pull trailing 90-day revenue. Take the top four or five that drive margin, not just units. These are the products a Brand Agent will be asked about most, and the ASINs Rufus will surface when someone shops your brand. If these cannot answer size, material, care, and the top comparison question in your category, nothing else in the sample matters yet.

One per failure shape (4–5 SKUs)

Fill the rest from problem types you already know:

• A bundle or kit whose components are not structured as child facts.
• A compatibility-heavy SKU (cases, filters, parts, replacements) where “works with” is prose.
• A gift or seasonal SKU with thin attributes and heavy lifestyle copy.
• A variant-heavy parent where size/color exist as options but weight, fill power, or thread count never became fields.
• An Amazon-only or dual-channel SKU whose backend attributes diverge from the Shopify metafields.

Ten is the ceiling for a free sample because the point is judgment density, not coverage. The playbook for large catalogs scales the same templates after the sample proves them.

1. 1

   Export the revenue cut

   Trailing 90 days, top SKUs by revenue. Note channel (Shopify, Amazon, both).

2. 2

   Tag each candidate with a failure shape

   Bundle, compatibility, gift/thin copy, variant-heavy, dual-channel drift — or “clean hero.”

3. 3

   Lock ten with at least four shapes represented

   Reject a sample that is ten clean heroes. Agents already handle those; they fail on the rest.

4. 4

   Attach source packs per SKU

   PDP URL, current metafield export or Amazon flat-file row, and any manual/spec PDF you already have.

What the enrichment pass actually does

For each of the ten, the work follows the same pipeline used on paid engagements — just narrower.

1. Inventory what a shopper (or agent) asks

We do not start from a generic attribute list. We start from questions: site search terms, support macros, return reasons, and the comparison questions shoppers type into ChatGPT or Perplexity for your category. Those questions cluster into candidate fields. Height and “will it fit under my desk?” are the same fact. That clustering is the seed of the product attribute schema.

2. Score the current record

For each SKU we mark every candidate field as: present and valid, present but wrong type or unit, present only in prose, or missing. Present-but-wrong is worse than missing — an agent will state a wrong metafield with confidence. Wrong beats empty in the bad direction, especially on compatibility data.

3. Extract, then verify

Prose extraction drafts values. Verification locks them. Manuals, supplier sheets, and measured packaging beat marketing adjectives. If a value cannot be verified, it stays blank with a callback note — we do not invent a fill-rate win.

4. Type the fields for the channel

Shopify gets metafield definitions with type, unit, and storefront access where agents need them. Amazon gets product-type attributes that match the flat-file or Category Specialist template for that ASIN. Dual-channel brands get a mapping table so the same fact does not diverge into two dialects.

That Amazon density is why a “we filled five metafields on Shopify” sample can still fail Vendor or Seller Central completeness checks. The sample is designed to surface that mismatch early.

5. Draft the category schema, not only ten rows

The durable output is the contract: required vs recommended fields, allowed values for enums, units for numbers, and the publishing gate. Ten filled rows without a schema means the eleventh SKU arrives half-empty again. Ten rows with a draft schema means you can estimate the full category honestly.

What you get back (the deliverable)

A useful sample package is small and concrete:

• Before/after table for each of the ten SKUs — field name, old value (or “prose only” / empty), new typed value, source of truth.
• Draft category schema — 8–20 fields typical for one category, tiered required / recommended / optional.
• Publish notes — Shopify metafield namespace and keys, or Amazon attribute column names; anything that needs a theme or A+ update called out separately.
• Gap list — facts that need a supplier callback or a warehouse measure before they can be filled.
• Go/no-go read — plain language on whether the category is enrichment-ready or blocked on missing source data.

You should be able to publish at least some of the ten the same week if your admin access is ready. If the package is only commentary with no pasteable fields, it failed as a sample.

How to read the result as a go/no-go

Use the sample as a decision tool, not a trophy.

Enrich the category next when most of the ten had facts available in prose, PDFs, or supplier sheets, and the draft schema is stable after one revision. That is Problem A. Scope the next 50–100 revenue SKUs in that category, then the siblings that share the same attribute pattern.

Pause agent configuration when more than about three of ten hero SKUs needed invented values or open supplier tickets for basic purchase-gating facts (fit, material, wattage, model list). Turning on a Brand Agent or pushing incomplete Amazon attributes will amplify wrong answers. Fix source data first.

Stay on metafields for now when the sample schema is under ~25 fields and one or two categories dominate revenue. A PIM is still optional; the schema and governance matter more than the tool. Revisit PIM only when multi-channel divergence and edit volume make Shopify-as-master painful — the same line we draw in broader catalog strategy.

Widen beyond one category only after the first schema ships with a publishing gate. Skipping the gate is how enriched catalogs decay within a quarter of normal product velocity.

Worked example: outdoor gear parent SKU

Imagine a mid-market outdoors brand. Hero tent SKU. Shopify description says “roomy for three, packs small, waterproof.” Amazon bullet list repeats the adjectives. Support tickets ask: floor area, peak height, packed size, season rating, and whether it fits a specific footprint.

Before enrichment the structured record has title, vendor, price, and three color variants. After the sample pass the same SKU carries typed fields: capacity_persons (3), floor_area_sqft (41), peak_height_in (48), packed_size_in (18x7), season_rating (3-season), minimum_footprint_sku (linked), fly_material (40D nylon ripstop), waterproof_rating_mm (1500). The draft category schema for “tents” marks capacity, floor area, peak height, and season as required — because those four questions dominate tickets and assistant prompts.

An agent can now answer “will this fit two adults and a dog?” with a field, not a guess. Rufus can compare peak height without scraping a paragraph. Off-site assistants that cite your PDP have corroborating structure instead of lifestyle prose alone. That is the entire point of the sample: one SKU makes the pattern obvious; ten SKUs make the project estimable.

Where this sits in the GigaCommerce stack

Catalog enrichment is the spine under agentic commerce and Commerce GEO. Brand Agents quote what you structured. Copilot Checkout still needs clean variants and policies, but it cannot invent product facts. Citation and recommendation work better when entities and attributes are consistent across channels.

If you are still deciding whether agents are even on the roadmap, run the readiness score first, then use the sample to prove the catalog half. If Amazon is the pain, send dual-channel SKUs in the ten so the attribute map includes Seller Central columns, not only Shopify keys. If you already know the catalog is thin, skip the debate and send the messiest revenue SKUs — that is what the free sample on Catalog Enrichment for AI is for.

Sibling reading in this cluster: the full enrichment playbook, attribute schema design, and compatibility attributes agents rely on.