Features
The core signals that decide whether AI agents can read, match, and surface your products.
Every AI shopping agent — ChatGPT, Perplexity, Rufus, Gemini — reads the same catalog fields before deciding who to surface. Legible measures the six core pillars every store is scored on — plus a seventh, multilingual coverage, if you sell in multiple languages — tells you where you're invisible, ranks the fixes by how much each one moves your score, and rewrites the descriptions you approve — one click to Apply, 30 days to Undo.
Pillar 01
Product descriptions,
scored by AI against AI's own rubric.
When a buyer asks an agent "find me a breathable linen shirt under $80 that ships to Paris," the agent doesn't browse your store — it reads your product text and decides in under a second whether you match the intent. Generic copy like "A comfortable linen shirt. Great for summer." returns a near-zero relevance score. Concrete copy — material, fit, use case, context — ranks.
Legible sends every product description to Claude Haiku with a scoring rubric built from the same evaluation criteria LLM-based shopping agents use internally. The rubric checks for: named materials, dimensional fit language, stated use cases, concrete sensory attributes, and the absence of filler phrases that inflate word count without adding signal. Each product gets a score from 0 to 100. You get a ranked list of the products whose rewrites would move your overall catalog score the most — so you're not guessing which page to fix first.
The score is not a measure of marketing quality. A description can be beautifully written for a human and still score low for an agent. That gap is exactly what Legible surfaces.
Pillar 02
Duplicate detection,
powered by embeddings — not exact match.
Catalog bloat is invisible from the admin but obvious to an agent. When a product appears four times under slightly different names — "Linen Shirt Sand," "Washed Linen Shirt (Natural)," "Summer Linen Shirt," "Linen Shirt Beige" — agents see a cluttered, low-confidence catalog and reduce the weight they give any single listing. A clean catalog with clear differentiation between variants gets surfaced more reliably.
Legible embeds every product using Voyage AI's voyage-3-lite model, then clusters products by semantic similarity. This catches duplicates that exact-match SKU deduplication misses entirely: re-releases, color aliases, seasonal relabels, and variants accidentally published as standalone products. Each cluster gets a confidence score and a recommended action: merge, differentiate, or archive.
The output is designed for your ops team as much as your content team — every cluster shows the products involved, the similarity score driving the recommendation, and a deep-link into Shopify admin so you can action the merge or differentiation directly.
Pillar 03
Alt text coverage,
for screen readers and multimodal agents alike.
Multimodal AI agents don't just read product descriptions — they process your images too. When your images have empty or missing alt text, a multimodal agent falls back to a generic caption or skips the image entirely. The visual context that might have confirmed "yes, this matches" is lost. Legible audits every product image across your catalog and reports alt-text coverage per product so you know exactly which images are missing captions.
Most stores have partial coverage — a few hundred images captioned, the rest empty. Legible surfaces the highest-impact gaps first: product images on your top-traffic pages, images used in search results, and images on products that are already close to passing on other pillars. Fix those, and you move the score without touching every image in the catalog.
The audit is also WCAG-aligned, so the same fixes that improve agent readiness improve accessibility for screen reader users. One investment, two payoffs.
Pillar 04
Metafields coverage,
in the namespaces agents actually query.
Agents don't read prose to find product attributes — they query structured fields. When a buyer asks "breathable shirts for men, natural fabric, ships from Europe," the agent maps those constraints onto metafield values: material, gender, origin. If those fields are empty, your product doesn't match — even if the description says exactly the right thing in plain text.
Legible audits whether each product carries the metafield attributes AI shopping systems read — material, audience, gender, origin, care, fabric weight, fit, and the Shopify product taxonomy fields — and surfaces the products with the lowest coverage so you can prioritize the fixes that move the score most.
This pillar is one of the highest-leverage to fix because it's the attribute layer agents query most. A catalog with great descriptions but empty metafields still fails structured queries — which make up the majority of agent-commerce traffic.
material field — 68 of 78 products missing
critical
audience field — partial coverage, inconsistent values
improve
gender field — well covered, 4 products missing
good
Pillar 05
Taxonomy consistency,
because agents trust catalogs that don't contradict themselves.
Product type drift happens quietly. "Tee," "T-shirt," "Tshirt," "Graphic Tee," "Short Sleeve Top" — five labels for the same category, spread across a catalog that grew organically. From the admin it looks fine. To an agent building a product graph, it looks like five different product types, each with a thin set of examples. That incoherence reduces confidence in all five.
Legible flags products whose product_type field is empty or missing the standard Shopify taxonomy mapping — the inputs agents need to place your products in the right category. When a fix would lift your score, Legible can generate and write the category/taxonomy attribute for you — alongside structured metafields and description rewrites — but only on a specific suggestion you click Apply, with a pre-apply diff and 30-day Undo. Every applied write is logged and independently reversible; you stay in control.
Taxonomy consistency also affects Shopify's own AI-powered search features and the Shopify product taxonomy used by Google Shopping and Meta Commerce. Fixing it once improves your presence across all three surfaces.
Pillar 06
Multilingual audit,
because an agent querying in French reads your French copy.
If you sell into France, Germany, or Japan, buyers querying AI agents in those languages will read your translated content — not your English fallback. When that content is placeholder text, incomplete translations, or machine-translated-and-forgotten copy, the agent sees a low-quality product listing and down-ranks you for that market. Legible checks every locale you've enabled in Shopify against every product, and reports coverage per pillar per market.
The multilingual audit answers "does a translation exist?" for every product title and description in every locale you've enabled in Shopify. Each finding is tied to a specific product and locale so you can action it directly in Shopify's translation editor or your chosen translation app. Translation-quality scoring (machine-translation detection, length-ratio checks, per-locale metafield coverage) is on the v2 roadmap.
The multilingual audit is only as useful as your translation pipeline — Legible tells you where the gaps are; how you fill them is up to you.
Scope
What Legible doesn't do — and why.
Targeted writes only — and only on your click. Legible writes back, only when you click Apply on a specific suggestion, AI-generated product descriptions and structured product metafields (a subtitle, a care guide, and standard taxonomy attributes). Every applied change is logged and independently undoable for 30 days. Alt text, images, customer data, and orders are never written — those stay in Shopify admin where you control them directly.
This is a deliberate boundary. Catalog data is load-bearing: a bad automated rewrite can hurt your SEO, break search indexing, and invalidate translations. So every write goes through a pre-apply diff, a hallucination scan (regulatory claims, numbers, brand names), a prompt-injection guard, and a per-plan daily cap. If Shopify's current value diverges from what we wrote before you click Undo, we refuse to silently overwrite — you get a 3-way diff to resolve manually.
Bulk apply, metafield writes, and taxonomy writes are live — each gated behind the same per-action review, pre-apply diff, hallucination scan, and 30-day Undo. Legible still never writes alt text or images, and never writes customer or order data. If a capability you need isn't live yet, email us (hamza@trylegible.com) — we'd rather tell you honestly what's shipped than oversell.