When merchants ask what to change so ChatGPT, Perplexity, or Google AI will recommend their product, the honest answer starts with a correction:
Those systems are not grading your hero photography or your tone of voice. They are consuming evidence they can compare: identity, attributes, variants, price, availability, and identifiers. If that evidence is missing, inconsistent, or buried in prose, the model has nothing reliable to quote—so it quietly picks a competitor whose catalog is easier to read.
This post is a Shopify-specific field guide: what actually shows up in the structured layer merchants control, and why "we fixed the description" rarely moves the needle.
The page is marketing. The catalog is evidence.
Your Online Store theme is built for humans: layout, sections, upsells, reviews widgets, and paragraphs of story. That still matters for conversion.
But when an AI assistant answers "which trail running shoe under $180 is in stock in EU 42?", it is not simulating a shopper scrolling your PDP. It is working from normalized product records: fields and feeds that can be joined, filtered, and compared across merchants.
Shopify gives you two parallel truths:
- What the buyer sees — Liquid, images, accordions, SEO copy.
- What systems read — Admin fields, variant rows, metafields, taxonomies, and the feeds/APIs you export to Google, Meta, marketplaces, and increasingly to AI syndication partners.
If (2) is thin, (1) cannot rescue you in AI answers—because the model is not asked to appreciate your brand story. It is asked to justify a recommendation with comparable facts.
Identity: what is this product?
These fields answer the most basic question an AI must get right: what object are we talking about?
title— The primary label. If it is vague ("Pro Trainer v2"), the model leans harder on other fields. Clear nouns beat clever names.product_typeand taxonomy — Shopify's product type and any channel taxonomy you map (for example Google product category) tell assistants whether this is a shoe, a watch band, or a supplement. Ambiguous types are a silent killer: the system may match the wrong intent entirely.vendorandtags— Useful for disambiguation and filtering, dangerous when abused. A wall of promotional tags reads as noise, not signal.handleand canonical URL — Stability matters for deduplication across refreshes and feeds.
If identity fields disagree with each other—product_type says apparel but your structured attributes describe electronics—the model receives conflicting evidence. Conflicting evidence reads as low confidence. Low confidence means omission.
Variants: the matrix has to be real
Most AI shopping questions implicitly include constraints: size, color, width, capacity, voltage. That maps directly to Shopify variants.
What systems want is boring and strict:
- Every sellable combination you claim should exist as a variant row (or an explicitly modeled equivalent).
- SKU and barcode (GTIN/UPC/EAN) where the channel expects them—identifiers are how assistants dedupe "same product" across merchants.
- Price and compare-at price in the currency the buyer context expects.
- Inventory policy and quantity that match reality—ghost availability is worse than "out of stock" because it trains distrust across the whole catalog.
"We described sizes in the body" is not a substitute. A model cannot reliably sort EU 42 vs US 9 from prose at scale. It needs structured options (Size, Width, Color) and consistent option naming across the catalog.
Commerce signals: can the assistant safely recommend a purchase?
Even perfect specs fail if the buyer cannot act on them. AI shopping surfaces increasingly weight:
- Current price and currency
- Availability (in stock, preorder, backorder rules)
- Shipping regions or fulfillment signals when exposed
- Return policy and warranty when available as structured or linked policy pages
These fields are where "technically correct in the CMS" still fails in the real world: stale inventory, variant-level price drift, or a default "ships worldwide" that is not true will contradict other feeds the same assistant ingests.
Metafields: where the real catalog actually lives
Out of the box, Shopify gives you a strong shell. The differentiation for serious merchants is almost always in metafields: material composition, care instructions, dimensions, compatibility matrices, certifications, ingredients, battery chemistry, water resistance, and hundreds of vertical-specific facts.
That is good news: you can model nuance without cluttering the title.
It is also where catalogs go wrong:
- Same concept stored under three different metafield keys across collections.
- Values as uncontrolled strings (
"blk","Black","#000"). - Variant-scoped facts stored at product level (or the reverse).
Metafields are only "structured" if they are consistent enough to compare. Otherwise they are just a second layer of prose.
Media and trust: what helps when it is structured
Images and reviews matter—but again, structure wins:
- Alt text and image order carry signal when assistants summarize "what it looks like."
- Review aggregates (rating, count) are easy evidence when exposed consistently—not screenshots of five stars in the hero.
If reviews only live in a third-party widget with no machine-readable aggregate, the assistant may behave as if reviews do not exist.
What body_html can and cannot do
Long descriptions help humans. For AI, prose is last-resort evidence: expensive to parse, easy to contradict, hard to compare across SKUs.
That does not mean you should delete story-driven copy. It means you should not rely on it as the only place where waterproofing, compatibility, or dimensions live.
If a fact matters for comparison, duplicate it into a structured field or metafield with a normalized value. The description can still explain why it matters; the field carries what is true.
So what should you do this week?
You do not need a philosophical debate about GEO to make progress. Open your best-selling SKU in Admin and answer:
- Is
product_typespecific enough that a stranger would classify it correctly in one try? - Are variants complete for every combination you sell—and are identifiers present where your channels require them?
- Are the top five comparison facts buyers ask for modeled as metafields, not buried in paragraphs?
- Do price and availability match what your feeds say to Google and other surfaces?
If any answer is "not sure," that is the work—not another round of "AI-optimized" adjectives.
How Listwiser fits (without pretending we read ChatGPT's weights)
Listwiser is built to make (2) the structured layer auditable: what is present, what is ambiguous, what contradicts another source, and what blocks AI-safe recommendations—per product and per variant, with provenance.
If you want a concrete picture of where your catalog sits today, run a read-only scan. No card, no rewrite-by-default—just the same structured truth assistants lean on, surfaced the way your team can actually fix it.
The acronyms will keep changing. Fields will not.