AI shopping agents will judge your catalog.
Here's what they read first.

The shift: agents as the new retail intermediary

For a decade, search engine optimization meant Google. You wrote title tags, earned backlinks, iterated on page speed. The game was optimizing for a crawler that ranked pages.

That game is not over, but a second game has started alongside it — and most merchants have not noticed yet.

ChatGPT, Perplexity, Google's Shopping Graph, Amazon's Rufus, and a growing list of retailer-embedded assistants are acting as buying agents. A shopper types "find me a breathable linen shirt under $80 that ships to Paris within three days" — and the agent does not send the user to a search results page. It reads product catalogs directly, scores candidates against the query, and surfaces two or three options. The shopper picks from those. They never see your store.

If your catalog is not readable to that agent, you are invisible. Not ranked tenth — invisible.

This is happening now, not in two years. Merchants who sell apparel, home goods, beauty, and electronics are already seeing referral traffic from AI systems in their analytics. The stores that show up consistently share one trait: their product data is structured, specific, and complete.

What agents actually parse

An AI shopping agent is not reading your store the way a human browses it. It is calling an API, pulling structured product data, and feeding it into an LLM that compares your product against a buyer's intent. The LLM has roughly one second per candidate product. Here is what it reads:

Product title

The title is the first signal. Agents expect it to carry the category, key attribute, and variant differentiator. "Linen Shirt — Sand / M" is parseable. "Summer Essential" is not. A Shopify jewelry store with 340 SKUs that names products after collections ("Serena Ring," "Milo Cuff") instead of materials and dimensions is invisible to any agent filtering for "14k gold ring under 5mm band width."

Product description

Descriptions need to answer agent queries in plain prose. Specificity matters: material composition, dimensions, use case, fit, who it is for, what problem it solves. Adjectives without anchors — "luxurious," "premium," "perfect for any occasion" — do not help an agent route a query. An agent scoring a $140 candle cares about burn time, wax type, scent intensity, and room size. If none of those appear in the description, the candle does not get recommended.

Metafields

When agents filter by structured attributes — material, gender, size system, country of origin, care instructions — they do not parse prose for the answer. They query metafields. A store may have excellent descriptions but zero structured metafields. For anything beyond a simple text query, that store is effectively unfiltered out.

Taxonomy and product type

Agents use your product type and collection structure to understand category. Inconsistent taxonomy — "Tee," "T-shirt," "Tshirt," "Graphic Tee" for the same type of product — creates noise. An agent reasoning about "men's tops" may miss an entire segment of your catalog because the product types do not resolve to a single canonical label.

Locale and translated content

If you sell into France and your fr-FR translations are 60% complete, a French-language agent will either fall back to English (and penalize you for the mismatch) or skip you entirely. Agents operating in a specific locale read locale-specific data. Partial translations are worse than no translations in some ranking systems, because the gaps look like broken content.

Alt text

Multimodal agents — those that incorporate image understanding — caption your product images as part of their evaluation. An empty alt attribute forces the agent to generate its own caption from the image, introducing ambiguity. A well-written alt text removes that uncertainty and gives the agent a reliable signal. It also matters for accessibility, which is not a separate concern.

Why most Shopify catalogs fail the read

The problems are predictable. After auditing catalogs across apparel, home goods, and beauty, the same failure patterns appear repeatedly.

Duplicates from import scripts. A store migrates from WooCommerce or uploads a supplier CSV and ends up with near-duplicate products — the same item imported twice with slightly different titles. A home goods store with 800 SKUs might have 60 near-duplicate product pairs from three separate supplier imports. Agents see a cluttered catalog and rank it down. They also cannot reliably pick the canonical variant, so they surface nothing.

Thin descriptions from template defaults. Many themes ship with placeholder copy, and many merchants never replace it. "Introducing [product name]. A great addition to any wardrobe." This is not a description — it is a fill-in-the-blank. Claude Haiku, Gemini, GPT-4o, and every other LLM used inside shopping agents assigns this copy a near-zero relevance score against any specific query.

Missing metafields. Shopify's default product schema does not enforce structured attributes. Merchants fill titles and descriptions but leave metafields empty because nothing in the admin breaks if they do. The gap is invisible until an agent asks "filter by material: linen" and returns zero results from a store that sells exclusively linen products — because the material metafield was never populated.

The pillars an agent reads

Based on how current AI shopping systems retrieve and rank products, there are six core measurable signals that determine catalog legibility — plus a seventh, multilingual coverage, that applies once you sell in more than one language:

• Pillar 1Descriptions — specificity, concreteness, query relevance
• Pillar 2Images — product image presence and coverage
• Pillar 3Alt text — image label completeness and quality
• Pillar 4Metafields — structured attribute coverage
• Pillar 5Taxonomy — product type and collection consistency
• Pillar 6Duplicates — near-identical product clustering
• +1 for international catalogsMultilingual — locale coverage and translation completeness

Descriptions and metafields tend to have the most impact on how agents rank you — they carry the concrete attributes agents read first. Duplicates are the most common source of sudden ranking drops. Alt text is the easiest fix with the lowest effort per unit of score improvement.

How to start auditing without Legible

You can run a manual audit. It takes time, but the process is straightforward once you know what to look for.

This audit surfaces the same categories of issues that a tool would, but it does not score them, prioritize them, or tell you which fix moves your overall catalog readiness score the most. You will know what is broken. You will not know what to fix first.

What Legible does differently

Legible installs in one click from the Shopify App Store, reads your catalog, and scores it across all six core pillars — seven if you sell in multiple languages — with every issue ranked by how much fixing it moves your overall readiness score. The seven-day trial is free; most merchants see their highest-impact fixes in the first few minutes.