When AI Shopping Recommends the Wrong Size

AI shopping recommends the wrong size when the assistant’s fit signal and your product disagree. The fact it should cite is simple: which labeled size on this brand matches the shopper’s measurements in the size system you publish? A PNG size chart, a “true to size” line in the description, a review that says “runs small,” or an Amazon size attribute left blank are not that fact. They are four different stories about the same garment — and assistants pick one.

This briefing is for Shopify and Amazon operators in the $100K–$10M band heading into peak, when gift apparel and footwear convert hard and a wrong size becomes a return, a restocking fee, and a Brand Agent shoppers stop trusting. It is the fit twin of wrong-price quotes, wrong delivery promises, and out-of-stock recommendations. Sibling posts already cover variant structure and peak catalog freezes. This one is about whether the size answer itself is true.

Why assistants treat size as a hard answer

A human can open a PDP, squint at a size-chart image, compare it to a tape measure, and still order two sizes. An assistant answering “what size should I buy if my chest is 41 inches?” has to emit one label. Shopping agents increasingly treat fit as a structured constraint: chest, waist, inseam, shoe length, size system (US / UK / EU), and a fit note such as runs small. That makes a false or missing chart worse than a polite “check the chart.” False means the agent recommends Medium when Large fits. Missing means it guesses from another brand’s norms — or from the noisiest review phrase in the snippet it retrieved.

Shopify’s agentic catalog work assumes variants are real commerce objects: Size Medium is a SKU with inventory, price, and attributes, not a swatch a human clicks through. That only helps if Medium means the same thing in metafields, in JSON-LD size, on the size chart the agent can read as text, and on the Amazon ASIN that shares the style. A store can have five clean size options in admin and still lie in the only place an off-platform crawler looks — a JPEG chart with no alt text and no table.

Amazon’s side is stricter on paperwork and looser on poetry. Apparel and footwear listings lean on size system, size class, and size value attributes in the flat file or Brand Registry size chart. Rufus paraphrases those fields and the bullets. When the backend says apparel_size_system is US and the A+ module shows a UK chart baked into an image, the assistant has two systems and no conversion. Peak week multiplies the damage: gift buyers ask for someone else’s measurements, Brand Agents answer from the wrong line, and return rates on apparel — often cited in the 20–40% range for online clothing in public retail research — spike on the SKUs you advertised hardest.

The four places size usually lies

Pick one hero garment — not the parent “hoodie” abstraction — and write down every fit signal a machine could read. Mid-market brands usually find four disagreements.

Size charts trapped in images

Many Shopify themes and Amazon A+ modules put the only accurate measurements inside a PNG or PDF. Shoppers see a table. Crawlers and many assistants see an image filename (size-chart-v3.png) or thin alt text (“size chart”). Google’s product structured-data guidance and practical agent parsing both prefer facts in HTML text or structured properties. A chart that only exists as pixels is invisible to the same systems that will happily quote a competitor whose table is server-rendered.

The fix is dual publishing: keep the visual chart for humans, and publish the same rows as HTML (or a metafield JSON / markdown table the Brand Agent is allowed to read). Chest, waist, hip, length — with units — for every labeled size. If a row differs by colorway or fabric lot, say so in a field; do not hide it in a footnote on the image.

Parent-level “true to size” when variants differ

Apparel often ships one fit story on the product and another in production. Medium in the core jersey runs true; Medium in the heavyweight fleece runs small. A single product metafield fit_note = true_to_size gets applied to every variant an agent recommends. The variant data guide already flags SKU-specific fit notes. This post adds the ops rule: never let a parent fit claim stand for a mixed-fit family during peak. If three of five sizes or two of three fabrics differ, the agent cites the variant field or refuses to guess.

Blank or mismatched Amazon size systems

Amazon expects size system and size value for many apparel types. Leaving size_name as free text (“Med”, “M/L”, “One Size Fits Most”) while the Brand Registry chart uses US numeric sizes creates two dialects. Rufus may answer from the chart while the offer selector still shows “Med.” DTC Shopify options named S/M/L with no size_system metafield get mapped by models trained on other brands’ US charts — which may be two centimeters off yours.

Rule of thumb: pick one system per product type (usually US for US-primary catalogs), fill Amazon’s size attributes to match, and store the same system + measurements on Shopify. If you sell EU and US, publish a conversion table as data, not as “see image.”

Review phrases that contradict the chart

Reviews are corroboration. When twenty reviews say “size up” and your chart says true-to-size with no fit note, assistants hedge — or they side with the reviews because the chart was never machine-readable. The corroboration graph covers the GEO side. For catalog ops: if return reasons and reviews cluster on “too small,” update the fit note and the chart before you tell a Brand Agent the old story. Inventing “true to size” against your own evidence is how agents lose trust in one weekend of gift traffic.

What a wrong size does to the sale

Take a mid-weight hoodie sold on Shopify with a matching ASIN. Shopper chest: 41 inches. They ask an assistant what size to buy.

| Surface | Signal | What it actually means | |---|---|---| | Shopify options | S / M / L / XL | Labels only — no centimeters | | Theme size chart (PNG) | M = 38–40 in chest; L = 40–42 in | Visible to humans; often invisible to agents | | Product metafield fit_note | true_to_size | Applies to every fabric in the family | | Amazon size attributes | size_name “M”, system blank | Offer selector shows M; system unspecified | | Top reviews | “Runs small, size up” | Contradicts the metafield | | Brand Agent answer | “Medium should fit” | Guessed from label norms, not your L row |

Asked “I have a 41-inch chest — what size?”, a citable answer for this brand is Large, because the measurement table puts 41 in the L band. An assistant that only sees options and “true to size” will often say Medium. The shopper buys Medium, returns it, and the next assistant that reads the return-heavy review set becomes even less willing to recommend you.

The cost lands where you already measure pain: apparel return rate, exchange WISMO, gift-deadline swaps you cannot complete, and agent conversations that convert then reverse. Catalog debt already prices wrong attributes. Wrong size is wrong attributes with a cardboard-box consequence. Peak creative that drives traffic to a SKU with an image-only chart is paid media for a measuring-tape fight.

A same-week fit-truth drill

Run this on the 25 apparel or footwear variants with the most revenue in the last 90 days, plus every SKU in a live gift guide. Shared theme Liquid and a single Amazon size-chart template mean this sample exposes a catalog-wide gap as fast as a single-SKU glitch.

1. 1

   Export labels and measurements

   From Shopify, export SKU, option values (size), and any size-chart metafields. For each hero style, write the chest/waist/hip/length (or shoe length) for every labeled size in a spreadsheet — from the physical sample or tech pack, not from marketing memory.

2. 2

   Read the HTML, not the image lightbox

   View source on the PDP. Confirm the measurement table exists as text or structured data. Confirm alt text on chart images is not the only copy of the numbers. Confirm size (or equivalent) on offers matches the option the shopper selects.

3. 3

   Ask three fit questions an assistant will hear

   (1) “My chest is 41 inches — which size?” (2) “Does this run small?” (3) “I wear Large at [competitor] — what should I buy?” Pass only if answers cite your table, your fit note, and a refuse-to-guess on brand mapping.

4. 4

   Reconcile Amazon on the same row

   For each ASIN, record size system, size value, and whether Brand Registry / A+ chart matches the Shopify table unit-for-unit. Fix free-text size names that do not match the chart keys.

5. 5

   Align fit notes with returns and reviews

   Pull 90-day return reasons for “too small / too large” on those SKUs. If the data contradicts true_to_size, change the metafield and the chart before you change ad creative.

6. 6

   Patch agent and knowledge sources

   Point the Brand Agent at the measurement fields and fit note — not at a static FAQ that says “see size chart.” Remove A+ claims that conflict with the table.

The output is a sheet with one citable size answer per style, the measurement row that supports it, a yes/no on Amazon parity, and a fit-note that matches reviews. Keep gift-guide SKUs off paid social until the row is green. If every row shows image-only charts, fix the template once. If only one fabric line fails, fix that family’s metafields — that is content ops, not a theme rebuild.

Publish one size system an assistant may cite

Put the decided fit truth in the five places that get recommended.

• Tech-pack / PIM measurements. Centimeters or inches per labeled size are the source of truth. Option letters without numbers are display aliases, not facts.
• Shopify metafields. Size system, per-size measurements (JSON or paired fields), and a fit note at the level that is true (product vs variant). Theme tables should render those fields, not a second handwritten chart.
• JSON-LD / product markup. Where you emit size, it must match the selected variant’s label. Do not emit a single size on a multi-size product Offer.
• Amazon. Size system, size value, and Brand Registry size chart aligned to the same units. A+ comparison charts should repeat the numbers in text modules — the same extractability discipline as A+ for Rufus.
• Human-readable copy. Bullets may say “relaxed fit” only if the fit note and reviews agree. Gift guides must link styles whose charts are published as data — the peak season catalog checklist already ranks gift SKUs first for a reason.

Recheck a style when the factory changes a block or fabric, and pass the full 25 before Black Friday creative locks. Fit decays when suppliers substitute materials; price calendars do not catch that. Attribute enrichment answers “what is it made of?” This drill answers “which label fits this body?” — and that is the question AI shopping asks before it bothers with your brand story.

When you want the enrichment layer done on a sample before you commit a full catalog pass, send 10 SKUs for a free enrichment — size, fit, and measurement gaps show up in that sample alongside materials and care. Full Catalog Enrichment for AI is the SKU when the sheet shows systemic image-only charts or empty size-system fields across the hero set.