Shopify product descriptions for AI visibility: structure, factuals, length

How AI agents parse descriptions (hierarchically)

An LLM doesn't read your description top-to-bottom the way a human does. It chunks the HTML, ranks chunks by perceived information density, then attends most strongly to the high-density blocks. The order of preference, consistently across the 6 agents we track:

• Bulleted lists (<ul><li>).High density, low ambiguity. Each list item is a self-contained fact.
• Subheadings (<h3>).Semantic anchors. They tell the agent "this section is about Material".
• Spec lines ("Material: 100% merino wool").Pseudo-structured even inside prose.
• Paragraph prose.Read last, weighted least. Most fluff lives here.

The implication: where you put a fact matters as much as whether the fact is present. Hiding "100% merino wool" inside a 200-word brand-story paragraph is functionally invisible. The same words in a bulleted list are far more reliably retrieved, without rewriting the content. (We don't have a clean controlled measurement of the exact weight difference, only consistent qualitative observation across our benchmark suite.)

The 7 criteria from the rubric

Source: commerce-agentic/agentic-catalog-scanner (CC0). Every criterion is implemented in the open audit engine.

Floor checks (apply before the criteria above):

• Word count under 50 → score capped at 6/20 regardless of structure.
• Lorem Ipsum or Shopify template placeholder present → capped at 4/20.

The default description structure that hits all 7 criteria

One paragraph of context, then a bulleted spec list, then a use-case paragraph. Under 250 words total. Every product, every vertical.

<p>{One paragraph: what the product is, in concrete terms.
   What it's made of or contains. Who it's for.}</p>

<h3>Specs</h3>
<ul>
  <li><b>Material:</b> {fabric/composition with %}</li>
  <li><b>Size:</b> {dimensions/sizing range}</li>
  <li><b>Weight:</b> {numeric with unit}</li>
  <li><b>Care:</b> {wash/care instructions}</li>
</ul>

<h3>Use</h3>
<p>{Ideal for {use-case 1}. Designed for {use-case 2}.
   Works with {compatibility}. Best when {condition}.}</p>

That template alone hits criteria #1 (≥150 words), #2 (bullets), #3 (h3), #6 (spec list with labels), and gives a natural home for criteria #4 (factuals) and #5 (use-cases). Criterion #7 (no fluff) is a discipline thing: bake it into your style guide and never use "premium" again.

Concrete examples per vertical

Same structure across verticals: lede paragraph → bulleted specs with labels → use-case paragraph. The labels (Material:, Active:, Allergens:) act as informal field names that agents extract as pseudo-structured data. This is the difference between "the agent reads the text" and "the agent reads data."

The 5 antipatterns to drop

How to bulk-rewrite descriptions in Shopify

1. Export your products to CSV (Products → bulk actions → Export). Sort by description word count ascending. The bottom of the list is your biggest opportunity.
2. Rewrite a sample of 5-10 products using the template above. Re-run our public audit to verify the description score increase.
3. For full-catalog rewrites, the Shopify app bulk-runs the same template via a Claude prompt under a per-criterion pre-flight check (won't ship a rewrite that doesn't improve scoreDesc).

One product at a time, the rewrite is ~10 minutes. For a 100-product catalog this is a long afternoon manually or about an hour with the app's Maximize cascade.

What this article doesn't claim

• We don't have controlled before/after data showing "rewriting descriptions to hit all 7 criteria increases AI mentions by X%". We can't run that experiment without merchant cooperation and a clean counterfactual.
• We do have the open rubric, and the consistent observation that catalogs hitting the 7 criteria predominate in our captures dataset.
• We do observe that catalogs scoring under 6/20 on descriptions are systematically absent from top recommendations. The floor checks exist because that pattern is real.

With titles (signal #1, 15 pts) and descriptions (signal #2, 20 pts) handled, the next highest-leverage move is the structured fields agents actively prefer: the metafields guide, vertical by vertical (signal #5, 15 pts).