How AI shopping agents read your products, and where they go blind. Independent research into the catalog an agent reads.
The finding
A shopper asks an AI agent for exactly what you sell. It hands them a competitor.
The agent read your listing. It just could not act on the one detail that made yours the right answer. Here is how a single request goes wrong:
1
A shopper describes what they want, in their words. A need and a budget, bundled into a few words. waterproof boots under $150
2
The catalog filters the price, but not the point. It can hard-filter the $150. It cannot filter “waterproof” at all, so it returns boots that soak through, as long as they are under budget.
3
The agent recommends from that list. Your waterproof boot competes against the ones the agent could not rule out.
To hard-filter is to keep only the products that carry a value and drop the rest, the way a size dropdown does. The left is the complete list an agent can do that on; the right is a sample of what it can only read in your text.
The agent can hard-filter · complete list
Price, as a budget
Star rating and review count
Price tier: budget, mid, premium
Condition: new or used
In stock now
Category
Colour — a dozen base names
Size and target gender
It can only read in text · a sample
Material
Occasion or use case
Fit and true-to-size
Waterproofing and performance
Heel height, sleeve, neckline
Gemstone and metal
Ingredient, SPF, skin concern
Scent and flavour
Anything ruled out (“not leather”)
Ask it to filter on waterproofing and it replies, word for word: Water resistance level is not supported and was ignored. Supported attributes: Color, Size, Target gender. So we pushed on the three it does allow, and even they are shakier than they look. Colour matches only Shopify’s base names, so a merchant’s own cinnabar dress returns nothing. Size values run from 10/M to L (fits EU 42–44), and the filter misses more than it lands. When stock runs thin, every filter quietly relaxes and lets the wrong products through. You can structure waterproof or occasion perfectly and still never filter on it. That ceiling is the platform’s, not yours.
Source: Atomz AI · The Catalog Visibility Report · 2026
52%Readrecovered from your product text
6%Filteredstructured so an agent can filter on it
56%Rule-outs keptstill fit when a shopper excludes one
AI shopping has stopped being a rounding error. Shopify reports agent-driven sessions up 8x and orders up 13x year over year, and agents now land on the product page in more than half of their sessions, against roughly a fifth for organic search. Shoppers who arrive through the catalog convert at twice the rate of other sources. The front door to your store is shifting from a person browsing to an agent querying.
When that shopper asks ChatGPT or another assistant to find a product, the assistant does not open your storefront. It queries Shopify's Catalog API, the global product catalog Shopify built as the discovery layer of the Universal Commerce Protocol, the open agentic-commerce standard it developed with Google and that Amazon, Mastercard, Meta, Microsoft, Stripe, Target, Walmart and Visa have backed. A single call searches across every Shopify merchant at once and returns a structured view of each product: title, description, price, media, variant options and availability.
Shopify does not just pass your fields through. Its machine-learning models read your titles, descriptions and images and infer the structured attributes wherever the facts are stated, turning plain prose into filterable data on your behalf. But the model can only extract what is written or shown. In Shopify's own words, “if that information is just missing altogether, then the attributes will also be missing,” and one missing field can drop a product out of an agent's results entirely. That inferred record, not your live page, is the surface where discovery now happens. This report reads it the way an agent does and measures exactly what it exposes and where it goes blind.
Two things an agent does with your data are not the same. It can read your text and match on meaning, and that mostly works. It cannot reliably filter: hard-narrow to only the products that qualify and drop the rest, the way a size or price dropdown does. This report is about that second gap.
1
Shopper · in ChatGPTA person asks in plain language.askscocktail dress under $100
2
ChatGPT · the agentIt turns that into a search and calls Shopify Catalog over UCP, the open agentic-commerce standard, never crawling your store.callssearch_catalog(query, country)
3
Shopify · CatalogIt answers from a standing, ML-built record of every product, not your live store. It infers attributes from your titles, descriptions and images, but cannot extract a fact that is absent from all three.
4
Shopify · responseIt returns a ranked set of products as that structured record.returnstech_specs · options · price · checkout eligibility
5
ChatGPT → ShopperIt reasons over only what came back, ranks, and answers. It cannot recover anything the record left out.showsone pick, and a checkout it hands off
Source: Atomz AI · The Catalog Visibility Report · 2026
Figure. One shopping request, end to end. The catalog record, not your storefront, is the surface the agent reads.
The agent reads almost everything, and filters almost nothing
Shopify's models recover most of what a shopper asks for from a product's plain text. So the exposure is narrow but sharp: the attributes stated nowhere, and the precise queries the agent cannot filter. This report reads the Shopify catalog the way an AI shopping agent does, across eleven categories, with every attribute grounded in Shopify's own Standard Product Taxonomy.
Readablerecovered from your listing text
Filterablestructured for a hard filter
Missingstated nowhere at all
Rule-outs keptstill fit when a shopper excludes one
The Catalog Visibility Score is the share of a product's demand attributes that are stated where an AI shopping agent can read them, averaged across the catalog. It measures how legible a catalog is to an agent, not sales or revenue. The structured, filterable layer, the forward-looking one, is a separate 6%.
Source: Atomz AI
How the 52 is built
For every product we list the things a shopper in that category asks about, then count how many are stated in text the agent can read. A boot buyer weighs five things: waterproofing, width, whether it resoles, the leather, and what it is for. A listing that states three of the five scores 60. Average that fraction across every product and the whole catalog lands at 52.
Electronics
34
Sports
39
Home & Garden
41
Apparel
46
Automotive
52
Food & Drink
60
Footwear
60
Pet
62
Cosmetics & skincare
65
Health & wellness
67
Jewellery
67
Figure. Catalog Visibility Score by category, the stated coverage of demand attributes. An index of legibility to an agent, not a revenue estimate.
01 · The market in one picture
Every category is readable, and none is actionable
Plot each of the eleven categories by how much of its demand the agent can read in text against how much it can filter on. All eleven land in the same quadrant: high on readable, flat on filterable. Readable is not the same as actionable, and today the catalog delivers only the first.
Figure 1. Seven categories by findable versus filterable, dimensional attributes weighted by products.
02 · Finding one
The structured layer is empty
For every applicable product we asked one question per attribute: is it a filterable option, merely findable in the text, or missing entirely.
Filterable
A structured option the agent can hard-filter on, like a real size or colour dropdown. The strongest state, and the rarest.
Findable
Written somewhere in the product text, so the agent can read it and match on it, but cannot filter by it.
Missing
Stated nowhere in the listing, so the agent cannot see it at all. This is the real gap.
Plot every demand attribute by how findable it is against how filterable it is, and the shape says it in one image: a dense band pinned to the floor. Almost everything is legible. Almost nothing can be filtered.
DescriptiveCompositionDimensionFitUse
Figure 2. 229 dimensional attributes, findable against filterable, coloured by role.
Why the filterable layer is empty. The 153 of 229 attributes that are filterable on 5% or fewer of products are not random. Everything merchants do structure is material, colour, size and flavour, the attributes Shopify's variant picker forces them to enter. They structure what the interface requires and nothing beyond it. Fit, waist rise, occasion, gemstone and protein source, the attributes precise queries depend on, are left to prose.
The pattern is checkable, brand by brand
It is not abstract. Here is one well-known footwear brand with a gap sitting on the exact attribute it is known for, and then the same thing playing out live in the catalog an agent reads.
Checkable example
Thursday Boot Company
Thursday built its name on rugged, weather-ready leather. Yet an agent can filter their boots on only two things, colour and size. There is no waterproofing field for it to filter on, so when a shopper asks for waterproof leather boots, the agent falls back on meaning and Thursday’s signature strength counts for nothing it can act on.
of the sixteen attributes Shopify defines for a boot are exposed as filters, colour and size
Run it yourselfwaterproof leather bootsmatched on meaning, never filtered
Live from the Shopify Catalog API · what an agent returns
A shopper searches waterproof leather boots
These are the four products the agent puts first. Thursday Boot Company, built on waterproof leather boots, is nowhere in them — there is no waterproofing field for the agent to filter on, so it ranks on meaning and Thursday competes on luck.
Group every attribute by what it describes, and the agent's reach drops sharply from left to right: it reads what a product is, and loses what it is for.
Descriptive
43% missing
Composition
33% missing
Dimension / numeric
55% missing
Fit / form
53% missing
Use / intent
44% missing
Source: Atomz AI · The Catalog Visibility Report · 2026
filterable (structured)findable (in text)missing (stated nowhere)
Figure 3. Three-state coverage by attribute role, weighted by products.
The agent reads what a product is. Material, colour, ingredient and design come back legibly, 60 to 90% of the time, whether or not the merchant structured them. It goes dark on how a product fits, how big it is, and what it is for: fit, waist rise, size and occasion are missing on roughly half of products, and structured on almost none. The same gradient holds in every category we measured, and the per-category detail is in the scorecards below.
Footwear
40% missing
Apparel
54% missing
Jewellery
33% missing
Cosmetics & skincare
35% missing
Pet
38% missing
Health & wellness
33% missing
Automotive
48% missing
Electronics
66% missing
Home & Garden
59% missing
Food & Drink
40% missing
Sports
61% missing
04 · Finding three
Precise queries are a coin flip on exclusions and price
Shopify describes the shift as the unit of shopping moving from a keyword to a use case: the agent decomposes a request into requirements, then matches each one against product specs. A requirement it cannot verify is a product it drops. Shopify's own example is a rain jacket that "must survive a torrential downpour" being ruled out when the listing never states a waterproof membrane. So legibility only matters if it changes what comes back. We put live queries through the catalog, typed by intent, and scored the top ten against every constraint. Positive and stated-numeric queries mostly resolve. Two things are unreliable, and both are the platform's, not the merchant's: the agent honours a ruled-out term only about half the time, and it does not hard-filter numeric ranges like price. It reads "not linen" and still returns linen; it reads "under $120" and still returns products over it.
We ran real shopper queries and, each time, checked the ten products the agent handed back. The score below is how many of those ten fit everything the shopper asked for, so 90% means nine of ten fit, 30% means only three did.
What the shopper asks for
How often the results fit
What happens
A specific number
moisturizer with SPF 30 or higher
92%
The number is written in the listing, so the agent respects it.
Several filters at once
waterproof leather chelsea boots
79%
Usually works. It only breaks when one of the filters is an attribute merchants leave out.
A price limit
leather boots under $120
39%
Price is structured, but the agent treats it as a hint, not a filter, so over-budget products come back.
Ruling something out
summer dress not made of linen
56%
The weak spot. Asked to rule linen out, the agent returned linen in nine of ten results. It works for some exclusions and fails for others; a ruled-out term is honoured about half the time.
Source: Atomz AI · The Catalog Visibility Report · 2026
Figure 7. Share of the agent's top-ten results that fit the query, by query type, averaged across six categories.
9 of 10 results the Shopify Catalog API returned for cocktail dress under $100 cost more than $100, even though cheaper cocktail dresses sit in the same catalog. The search matched the words and read “under $100” as a hint, not a filter.
The bridge to the supply side. Where a compound query fails, the cause is a missing attribute. "Long sleeve satin wedding-guest dress" scored 10%, because occasion is missing on 73% of dresses. The gap in the catalog and the failure in the query are the same thing.
05 · The one signal that is structured
Reviews are everywhere, and they still barely help
Attributes are missing, and subscriptions and loyalty are invisible, but one demand signal the catalog does expose on almost every product is the review rating: a star value and a review count. Across the products we sampled through the same search surface an agent uses, 85% carry one, against the 6% that carry a filterable attribute. It is the richest structured signal an agent has to rank on.
Category
Share of products with a review score
Median reviews
Median stars
Pet
97%
400
4.8
Jewellery
96%
337
4.8
Food & Drink
95%
616
4.9
Health & wellness
94%
1,389
4.8
Footwear
93%
678
4.8
Cosmetics & skincare
92%
1,617
4.7
Sports
88%
160
4.8
Apparel
78%
426
4.8
Home & Garden
78%
215
4.8
Electronics
76%
268
4.8
Automotive
55%
34
4.9
Source: Atomz AI · The Catalog Visibility Report · 2026
Figure 6. Review coverage and volume by category, products sampled through the catalog search an agent uses.
Why it barely helps. The star value is saturated: the median product scores between 4.7 and 4.9 out of five in every category, so if an agent sorts by rating almost everything ties. The real tiebreaker is review volume, and that ranges from a median of 34 in automotive to 1,617 in cosmetics & skincare. The agent ranks on how many reviews a product has, not how good it is, which quietly rewards the incumbents who have been collecting reviews the longest. And the reviews that most sway an agent increasingly sit off the catalog entirely, in the community consensus of Reddit, YouTube and review sites that Shopify itself names as a top input and that no merchant controls.
06 · Category reference
Every category, to the leaf
The three findings hold in every vertical. Each card below is one category at a glance: its visibility score and its single biggest gap. Open any card for the full breakdown, subcategory by subcategory, with the live queries the agent handled. Each leaf is judged on the attributes Shopify's own Standard Product Taxonomy defines for it.
FootwearBiggest gap: Colour, missing from 59% of listingsvisibility
The category, shown on one boot
Across boots in this category, an agent can filter 0 of the 5 attributes Shopify defines for them. It reads 4 in the text but cannot filter on them and is blind to 1.
Source: Atomz AI · The Catalog Visibility Report · 2026
We ran real searches in footwear and graded the agent’s results
For each search we looked at the ten products the agent handed back and checked every one against every condition in the search. A product counts as a match only if it satisfies all of them. So 10 of 10 means the agent got every result right; 5 of 10 means five of its results broke a condition. The count stays high when the terms are ones the agent reads well. It falls when a term is one the catalog can neither pin down in the text nor filter on, like a specific material or an attribute the shopper ruled out.
matched every condition broke a conditionhighlighted the agent can hard-filter; the rest it can only match as text
9 of 10 results matchedmulti-filter
waterproof leather chelsea boots
9 of 10 results fit; the other one missed waterproof.
8 of 10 results matchedmulti-filter
pointed toe black stiletto heels
8 of 10 results fit; the other 2 missed pointed toe.
7 of 10 results matchedmulti-filter
open toe leather slide sandals
7 of 10 results fit; the other 3 missed open toe and leather.
7 of 10 results matchedmulti-filter
slip-on white leather sneakers
7 of 10 results fit; the other 3 missed slip-on and leather.
4 of 10 results matchedmulti-filter
suede penny loafers with a leather sole
4 of 10 results fit; the other 6 missed leather sole.
3 of 10 results matchedrule-out
waterproof boots not made of leather
the shopper asked for no leather; only 3 of 10 results honoured it.
How to rank better in footwear
The typical reality
In footwear, the agent can filter on little more than colour and size. Colour is missing on 59% of products.
What to do
State Shoe size, Colour and Toe style. The platform already reads Footwear material and Material for you, so there is no need to structure those.
ApparelBiggest gap: Neckline, missing from 64% of listingsvisibility
The category, shown on one dress
Across dresses in this category, an agent can filter 0 of the 4 attributes Shopify defines for them. It reads 2 in the text but cannot filter on them and is blind to 2.
Source: Atomz AI · The Catalog Visibility Report · 2026
We ran real searches in apparel and graded the agent’s results
For each search we looked at the ten products the agent handed back and checked every one against every condition in the search. A product counts as a match only if it satisfies all of them. So 10 of 10 means the agent got every result right; 5 of 10 means five of its results broke a condition. The count stays high when the terms are ones the agent reads well. It falls when a term is one the catalog can neither pin down in the text nor filter on, like a specific material or an attribute the shopper ruled out.
matched every condition broke a conditionhighlighted the agent can hard-filter; the rest it can only match as text
10 of 10 results matchedmulti-filter
v-neck long sleeve fitted top
All 10 results fit the request.
7 of 10 results matchedmulti-filter
long sleeve satin wedding guest dress
7 of 10 results fit; the other 3 missed long sleeve and satin.
6 of 10 results matchedprice cap
high waisted wide leg jeans under $90
6 of 10 results fit; the other 4 missed one filter.
6 of 10 results matchedmulti-filter
fleece lined waterproof winter jacket
6 of 10 results fit; the other 4 missed fleece lined and waterproof.
6 of 10 results matchedmulti-filter
high impact wireless sports bra
6 of 10 results fit; the other 4 missed high impact.
1 of 10 results matchedrule-out
summer trousers not made of linen
the shopper asked for no linen; only 1 of 10 results honoured it.
How to rank better in apparel
The typical reality
In apparel, the agent can filter on little more than colour. Neckline is missing on 64% of products.
What to do
State Lining type, Waist rise and Bra coverage. The platform already reads Material and Fabric for you, so there is no need to structure those.
JewelleryBiggest gap: Colour, missing from 41% of listingsvisibility
The category, shown on one ring
Across rings in this category, an agent can filter 0 of the 4 attributes Shopify defines for them. It reads 3 in the text but cannot filter on them and is blind to 1.
Ring sizeMissingJewelry materialReadableGemstone typeReadableRing designReadable
Lab Grown Diamond Solitaire Ring · linjer. Product image via Shopify's public catalog; it belongs to its owner.
Jewellery: the taxonomy spread, subcategory by subcategory
filterablereadable, not filterablemissing
Rings68% stated
Jewelry materialGemstone typeRing designRing size
Necklaces & pendants68% stated
ColourJewelry materialNecklace designNecklace length type
Earrings65% stated
ColourJewelry materialEarring designEarring closure type
Bracelets & bangles75% stated
ColourJewelry materialBracelet design
Charms53% stated
ColourJewelry materialGemstone type
Watches57% stated
Case materialWatch movementColour
Source: Atomz AI · The Catalog Visibility Report · 2026
We ran real searches in jewellery and graded the agent’s results
For each search we looked at the ten products the agent handed back and checked every one against every condition in the search. A product counts as a match only if it satisfies all of them. So 10 of 10 means the agent got every result right; 5 of 10 means five of its results broke a condition. The count stays high when the terms are ones the agent reads well. It falls when a term is one the catalog can neither pin down in the text nor filter on, like a specific material or an attribute the shopper ruled out.
matched every condition broke a conditionhighlighted the agent can hard-filter; the rest it can only match as text
10 of 10 results matchedmulti-filter
moissanite yellow gold engagement ring
All 10 results fit the request.
10 of 10 results matchedmulti-filter
gold hoop earrings
All 10 results fit the request.
10 of 10 results matchedmulti-filter
automatic stainless steel dive watch
All 10 results fit the request.
9 of 10 results matchedmulti-filter
gold pendant necklace with a chain
9 of 10 results fit; the other one missed gold.
5 of 10 results matchedrule-out
gold band ring without any diamonds
the shopper asked for no diamond; only 5 of 10 results honoured it.
How to rank better in jewellery
The typical reality
In jewellery, the agent can filter on little more than colour.
What to do
State Earring closure type, Ring size and Necklace length type. The platform already reads Jewelry material, Earring design and Bracelet design for you, so there is no need to structure those.
Cosmetics & skincareBiggest gap: Product form, missing from 35% of listingsvisibility
The category, shown on one serum
Across serums in this category, an agent can filter 0 of the 4 attributes Shopify defines for them. It reads 3 in the text but cannot filter on them and is blind to 1.
Suitable for skin typeReadableActive ingredientReadableSkin care effectReadableSPF levelMissing
Freshly Juiced Vitamin C Drop · soko-glam. Product image via Shopify's public catalog; it belongs to its owner.
Cosmetics & skincare: the taxonomy spread, subcategory by subcategory
Suitable for skin typeActive ingredientSkin care effectSPF level
Lashes & brows64% stated
ColourMaterial
Haircare47% stated
Active ingredientSuitable for hair typeProduct form
Fragrance50% stated
Olfactory familyProduct formOccasion
Bath & body61% stated
Suitable for skin typeActive ingredientProduct form
Tools & devices79% stated
Bristle material
Source: Atomz AI · The Catalog Visibility Report · 2026
We ran real searches in cosmetics & skincare and graded the agent’s results
For each search we looked at the ten products the agent handed back and checked every one against every condition in the search. A product counts as a match only if it satisfies all of them. So 10 of 10 means the agent got every result right; 5 of 10 means five of its results broke a condition. The count stays high when the terms are ones the agent reads well. It falls when a term is one the catalog can neither pin down in the text nor filter on, like a specific material or an attribute the shopper ruled out.
matched every condition broke a conditionhighlighted the agent can hard-filter; the rest it can only match as text
10 of 10 results matcheda number
spf 30 or higher facial moisturizer
All 10 results fit the request.
10 of 10 results matchedrule-out
fragrance-free vitamin c serum
the shopper asked for vitamin c; only 10 of 10 results honoured it.
10 of 10 results matchedmulti-filter
sulfate-free shampoo for curly hair
All 10 results fit the request.
10 of 10 results matchedrule-out
aluminum-free natural deodorant
the shopper asked for deodorant; only 10 of 10 results honoured it.
8 of 10 results matchedmulti-filter
woody amber eau de parfum
8 of 10 results fit; the other 2 missed edp.
7 of 10 results matchedmulti-filter
full coverage matte liquid foundation
7 of 10 results fit; the other 3 missed full coverage.
How to rank better in cosmetics & skincare
The typical reality
In cosmetics & skincare, the agent can filter on little more than colour.
What to do
State Occasion. The platform already reads Skin care effect, Bristle material and Olfactory family for you, so there is no need to structure those.
PetBiggest gap: Animal type, missing from 66% of listingsvisibility
The category, shown on one bag
Across bags in this category, an agent can filter 0 of the 4 attributes Shopify defines for them. It reads 3 in the text but cannot filter on them and is blind to 1.
Pet dietary requirementsReadableLife stageMissingPet food flavourReadablePet food formReadable
Grass-Fed Beef Grain-Free Dog Kibble · openfarm. Product image via Shopify's public catalog; it belongs to its owner.
Pet: the taxonomy spread, subcategory by subcategory
filterablereadable, not filterablemissing
Food & treats66% stated
Pet dietary requirementsPet food flavourPet food formLife stage
Supplements74% stated
Pet supply product formPet food supplementsPet food flavour
Toys54% stated
Toy/Game materialAccessory size
Grooming & health49% stated
IngredientsAnimal typeProduct form
Gear & accessories52% stated
ColourMaterialAccessory sizeAnimal type
Source: Atomz AI · The Catalog Visibility Report · 2026
We ran real searches in pet and graded the agent’s results
For each search we looked at the ten products the agent handed back and checked every one against every condition in the search. A product counts as a match only if it satisfies all of them. So 10 of 10 means the agent got every result right; 5 of 10 means five of its results broke a condition. The count stays high when the terms are ones the agent reads well. It falls when a term is one the catalog can neither pin down in the text nor filter on, like a specific material or an attribute the shopper ruled out.
matched every condition broke a conditionhighlighted the agent can hard-filter; the rest it can only match as text
9 of 10 results matchedmulti-filter
durable rubber chew toy for aggressive chewers
9 of 10 results fit; the other one missed durable.
6 of 10 results matchedmulti-filter
grain-free salmon dog food for puppies
6 of 10 results fit; the other 4 missed puppy.
6 of 10 results matchedrule-out
dry dog food without chicken
the shopper asked for no chicken; only 6 of 10 results honoured it.
4 of 10 results matchedmulti-filter
hip and joint soft chews for senior dogs
4 of 10 results fit; the other 6 missed senior.
How to rank better in pet
The typical reality
In pet, the agent can filter on little more than colour. Animal type is missing on 66% of products.
What to do
State Animal type, Accessory size and Life stage. The platform already reads Pet dietary requirements, Pet supply product form and Pet food supplements for you, so there is no need to structure those.
Health & wellnessBiggest gap: Flavour, missing from 49% of listingsvisibility
The category, shown on one supplement
Across supplements in this category, an agent can filter 0 of the 4 attributes Shopify defines for them. It reads 4 in the text but cannot filter on them.
Food supplement formReadableSupplement health focusReadableDietary preferencesReadableIngredient categoryReadable
Pure Encapsulations Vitamin D3 125 mcg (5,000 IU) · supplement-first. Product image via Shopify's public catalog; it belongs to its owner.
Health & wellness: the taxonomy spread, subcategory by subcategory
filterablereadable, not filterablemissing
Supplements & vitamins78% stated
Food supplement formSupplement health focusDietary preferencesIngredient category
Source: Atomz AI · The Catalog Visibility Report · 2026
We ran real searches in health & wellness and graded the agent’s results
For each search we looked at the ten products the agent handed back and checked every one against every condition in the search. A product counts as a match only if it satisfies all of them. So 10 of 10 means the agent got every result right; 5 of 10 means five of its results broke a condition. The count stays high when the terms are ones the agent reads well. It falls when a term is one the catalog can neither pin down in the text nor filter on, like a specific material or an attribute the shopper ruled out.
matched every condition broke a conditionhighlighted the agent can hard-filter; the rest it can only match as text
10 of 10 results matchedmulti-filter
vegan magnesium glycinate capsules
All 10 results fit the request.
10 of 10 results matchedrule-out
fluoride-free mint toothpaste
the shopper asked for toothpaste; only 10 of 10 results honoured it.
10 of 10 results matchedmulti-filter
adjustable knee compression brace
All 10 results fit the request.
8 of 10 results matcheda number
vitamin d 5000 iu softgels
8 of 10 results fit; the other 2 missed one filter.
4 of 10 results matchedrule-out
gluten-free protein bars without soy
the shopper asked for no soy; only 4 of 10 results honoured it.
How to rank better in health & wellness
The typical reality
In health & wellness, the agent can filter on little more than colour and size.
What to do
State Protein source, Compression level and Size. The platform already reads Food supplement form, Supplement health focus and Ingredient category for you, so there is no need to structure those.
AutomotiveBiggest gap: Manufacturer type, missing from 79% of listingsvisibility
The category, shown on one brake set
Across brake sets in this category, an agent can filter 0 of the 4 attributes Shopify defines for them. It reads 2 in the text but cannot filter on them and is blind to 2.
POWERSTOP Z26 Extreme Carbon Ceramic Brake Pads / 2012+ Jeep GC SRT & Durango SRT / Dodge Cars / 6-Piston Front & 4-Piston Rear Setup · kow-performance. Product image via Shopify's public catalog; it belongs to its owner.
Automotive: the taxonomy spread, subcategory by subcategory
filterablereadable, not filterablemissing
Brakes48% stated
FitmentItem materialBrake positionManufacturer type
Filters53% stated
FitmentItem materialManufacturer type
Lighting50% stated
ColourFitmentLight sourceManufacturer type
Engine & drivetrain44% stated
FitmentItem materialManufacturer typePosition
Suspension & steering50% stated
FitmentPositionManufacturer type
Exterior & body74% stated
ColourFitmentItem material
Wheels & tires50% stated
FitmentItem materialManufacturer type
Electrical & battery35% stated
FitmentManufacturer typeItem material
Interior & accessories68% stated
ColourFitmentItem material
Fluids & chemicals36% stated
Vehicle typeViscosityManufacturer type
Car care & wash43% stated
Product formVehicle application area
Source: Atomz AI · The Catalog Visibility Report · 2026
How to rank better in automotive
The typical reality
In automotive, fitment, the attribute that decides whether a part fits your car, is readable in the text 72% of the time but filterable on just 3%. An agent can tell that a brake pad mentions Civics; it cannot filter to the parts that fit a 2018 Civic.
What to do
Make fitment structured: the exact make, model and year a part fits, plus OEM or aftermarket. Today it lives in prose the agent can read but not filter on.
ElectronicsBiggest gap: Power source, missing from 74% of listingsvisibility
The category, shown on one laptop
Across laptops in this category, an agent can filter 0 of the 4 attributes Shopify defines for them. It is blind to 4.
Processor familyMissingMemory technologyMissingStorage drive type installedMissingOperating systemMissing
MacBook Pro - 14-inch - M5 Pro - 15core CPU - 16core GPU - 24GB - 1TB SSD - Silver · tigertech-shop. Product image via Shopify's public catalog; it belongs to its owner.
Electronics: the taxonomy spread, subcategory by subcategory
filterablereadable, not filterablemissing
Audio & headphones38% stated
Connectivity technologyPower sourceMaterialColour
Computers & laptops19% stated
Processor familyMemory technologyStorage drive type installedOperating system
Source: Atomz AI · The Catalog Visibility Report · 2026
We ran real searches in electronics and graded the agent’s results
For each search we looked at the ten products the agent handed back and checked every one against every condition in the search. A product counts as a match only if it satisfies all of them. So 10 of 10 means the agent got every result right; 5 of 10 means five of its results broke a condition. The count stays high when the terms are ones the agent reads well. It falls when a term is one the catalog can neither pin down in the text nor filter on, like a specific material or an attribute the shopper ruled out.
matched every condition broke a conditionhighlighted the agent can hard-filter; the rest it can only match as text
10 of 10 results matchedmulti-filter
4k oled smart tv
All 10 results fit the request.
10 of 10 results matchedmulti-filter
usb-c fast charging cable
All 10 results fit the request.
10 of 10 results matchedmulti-filter
wireless gaming controller
All 10 results fit the request.
9 of 10 results matchedmulti-filter
wireless noise cancelling over-ear headphones
9 of 10 results fit; the other one missed noise cancelling.
6 of 10 results matchedmulti-filter
gaming laptop with 16gb ram
6 of 10 results fit; the other 4 missed gaming and laptop.
1 of 10 results matchedprice cap
wireless charger under $30
1 of 10 results fit; the other 9 missed one filter.
How to rank better in electronics
The typical reality
In electronics, the agent can filter on little more than connector gender. Power source is missing on 74% of products.
What to do
State Compatible resolution, Display technology and Operating system. The platform already reads the descriptive fields for you, so there is no need to structure those.
Home & GardenBiggest gap: Pattern, missing from 63% of listingsvisibility
The category, shown on one chair
Across chairs in this category, an agent can filter 0 of the 4 attributes Shopify defines for them. It reads 1 in the text but cannot filter on them and is blind to 3.
Light colourLight temperatureBulb sizeEnergy efficiency class
Decor46% stated
Decoration materialColourPattern
Appliances28% stated
Energy efficiency classPower sourceColour
Source: Atomz AI · The Catalog Visibility Report · 2026
We ran real searches in home & garden and graded the agent’s results
For each search we looked at the ten products the agent handed back and checked every one against every condition in the search. A product counts as a match only if it satisfies all of them. So 10 of 10 means the agent got every result right; 5 of 10 means five of its results broke a condition. The count stays high when the terms are ones the agent reads well. It falls when a term is one the catalog can neither pin down in the text nor filter on, like a specific material or an attribute the shopper ruled out.
matched every condition broke a conditionhighlighted the agent can hard-filter; the rest it can only match as text
10 of 10 results matchedmulti-filter
solid wood dining table
All 10 results fit the request.
10 of 10 results matchedmulti-filter
cast iron non-stick frying pan
All 10 results fit the request.
10 of 10 results matchedmulti-filter
dimmable led floor lamp
All 10 results fit the request.
7 of 10 results matchedmulti-filter
cotton king size duvet cover
7 of 10 results fit; the other 3 missed cotton and king.
5 of 10 results matchedprice cap
stainless steel air fryer under $100
5 of 10 results fit; the other 5 missed air fryer.
3 of 10 results matchedrule-out
wall art without a frame
the shopper asked for no frame; only 4 of 10 results honoured it.
How to rank better in home & garden
The typical reality
In home & garden, the agent can filter on little more than colour and size. Pattern is missing on 63% of products.
What to do
State Bulb size, Energy efficiency class and Light colour. The platform already reads Fabric and Material for you, so there is no need to structure those.
Food & DrinkBiggest gap: Dietary preferences, missing from 46% of listingsvisibility
The category, shown on one snack
Across snacks in this category, an agent can filter 0 of the 4 attributes Shopify defines for them. It reads 3 in the text but cannot filter on them and is blind to 1.
SAKARA Plant-Based High-Protein Granola Mix - Nutrient-Rich Ingredients, Vegan, Gluten-Free, Non-GMO, No Sugar Granola, Supports Satiation and Boosts · j1mtiw-1t. Product image via Shopify's public catalog; it belongs to its owner.
Food & Drink: the taxonomy spread, subcategory by subcategory
filterablereadable, not filterablemissing
Beverages54% stated
FlavourDietary preferencesAllergen information
Coffee & tea58% stated
FlavourDietary preferences
Snacks52% stated
FlavourDietary preferencesAllergen informationFood product form
Candy & chocolate69% stated
FlavourAllergen informationDietary preferences
Pantry & food59% stated
FlavourDietary preferencesAllergen informationFood product form
Source: Atomz AI · The Catalog Visibility Report · 2026
We ran real searches in food & drink and graded the agent’s results
For each search we looked at the ten products the agent handed back and checked every one against every condition in the search. A product counts as a match only if it satisfies all of them. So 10 of 10 means the agent got every result right; 5 of 10 means five of its results broke a condition. The count stays high when the terms are ones the agent reads well. It falls when a term is one the catalog can neither pin down in the text nor filter on, like a specific material or an attribute the shopper ruled out.
matched every condition broke a conditionhighlighted the agent can hard-filter; the rest it can only match as text
10 of 10 results matchedmulti-filter
organic dark roast whole bean coffee
All 10 results fit the request.
10 of 10 results matchedmulti-filter
gluten free protein granola bars
All 10 results fit the request.
9 of 10 results matchedrule-out
sparkling water with no sugar
the shopper asked for no sugar; only 9 of 10 results honoured it.
8 of 10 results matchedmulti-filter
dark chocolate with sea salt
8 of 10 results fit; the other 2 missed dark chocolate.
1 of 10 results matchedprice cap
organic extra virgin olive oil under $20
1 of 10 results fit; the other 9 missed one filter.
How to rank better in food & drink
The typical reality
In food & drink, the agent can filter on little more than colour and size.
What to do
State Food product form. The platform already reads Flavour for you, so there is no need to structure those.
SportsBiggest gap: Pattern, missing from 73% of listingsvisibility
The category, shown on one dumbbell
Across dumbbells in this category, an agent can filter 1 of the 3 attributes Shopify defines for them. It reads 1 in the text but cannot filter on them and is blind to 1.
Snode AD80 Drop-proof Cast Iron Adjustable Dumbbells 80 LB · snodesport. Product image via Shopify's public catalog; it belongs to its owner.
Sports: the taxonomy spread, subcategory by subcategory
filterablereadable, not filterablemissing
Fitness equipment44% stated
ColourExercise equipment materialResistance level
Outdoor recreation51% stated
MaterialActivityColour
Cycling30% stated
ColourPattern
Team & racquet sports39% stated
ColourPattern
Water & winter sports30% stated
ColourPattern
Source: Atomz AI · The Catalog Visibility Report · 2026
We ran real searches in sports and graded the agent’s results
For each search we looked at the ten products the agent handed back and checked every one against every condition in the search. A product counts as a match only if it satisfies all of them. So 10 of 10 means the agent got every result right; 5 of 10 means five of its results broke a condition. The count stays high when the terms are ones the agent reads well. It falls when a term is one the catalog can neither pin down in the text nor filter on, like a specific material or an attribute the shopper ruled out.
matched every condition broke a conditionhighlighted the agent can hard-filter; the rest it can only match as text
10 of 10 results matchedmulti-filter
waterproof 2 person camping tent
All 10 results fit the request.
8 of 10 results matchedmulti-filter
adjustable dumbbell set
8 of 10 results fit; the other 2 missed adjustable.
8 of 10 results matchedmulti-filter
inflatable stand up paddle board
8 of 10 results fit; the other 2 missed inflatable and paddle board.
6 of 10 results matchedmulti-filter
lightweight road bike helmet
6 of 10 results fit; the other 4 missed helmet.
0 of 10 results matchedmulti-filter
professional carbon tennis racket
0 of 10 results fit; the other 10 missed tennis and racket.
How to rank better in sports
The typical reality
In sports, the agent can filter on little more than colour and size. Pattern is missing on 73% of products.
What to do
State Resistance level, Pattern and Activity. The platform already reads Material for you, so there is no need to structure those.
07 · What changes next
The platform ceiling
Two of the failures we measured, exclusions and hard numeric filtering, sit largely above the merchant. The agent matches on meaning rather than filtering, so it honours a ruled-out term only about half the time and treats a price ceiling as a hint, returning over-budget products anyway. That is exactly the job structured, filterable attributes exist to do. The moment Shopify begins enforcing its structured layer as real facets, the 6%-filterable number becomes the number that decides who is in the results. The merchants who structured early will be the ones who can be filtered to. Until then, the levers are the two below.
The callBy 2027, we expect filterability, not price or reviews, to become the attribute that separates the catalogs an agent can act on from those it cannot. The merchants who structure now are the ones an agent will be able to filter to the day the platform starts enforcing it.
The catalog is not the agent's only input. Shopify's own model of the agent weighs it alongside signals a merchant can only influence or adapt to: the community consensus of reviews and forums, search ranking on Google and Bing, and the agent's memory of the shopper. The catalog is the one layer you fully control, and getting into it cleanly is what earns a seat at that table.
Fill the gaps
State the dimensional attributes your category leaves blank: waist rise, occasion, protein source, life stage, fit. The model can only extract what the listing states, and one missing field can drop you from the results entirely.
Enrich the prose
The model reads text and images and infers structure from them, so richer, specific copy does more for recall today than empty metafields do. Shopify finds AI-discoverable shops carry 37% longer descriptions and half the empty fields of the rest.
Consolidate duplicates
Duplicate listings dilute the signal and make an agent pick arbitrarily between near-identical products. Use combined listings, or the Catalog mapping tool, to steer how your products cluster.
Skip the tag gymnastics
Tags are not handed to the agent; they only feed Shopify's inference. Plain-English facts in your copy matter more than tag structure. Spend the effort on descriptions, not taxonomy strings.
Glossary
What each term means
AI shopping agent
An assistant such as ChatGPT or Copilot that searches, compares and, increasingly, checks out on a shopper's behalf, instead of the shopper browsing a store themselves.
Shopify Catalog API (Global Catalog)
The machine-readable feed of Shopify's global product catalog that AI agents query to find products. When an assistant shops, this is the surface it reads, not your storefront.
Universal Commerce Protocol (UCP)
The open agentic-commerce standard Shopify developed with Google (and backed by Amazon, Mastercard, Meta, Stripe, Visa and others) that defines how agents search, cart and check out. The Catalog API is its discovery layer.
Standard Product Taxonomy
Shopify's published set of categories and the attributes (metafields) each one should carry. Every attribute measured in this report is one Shopify itself defines.
Attribute / metafield
A single structured product fact, such as material, heel height or fit. A metafield is where Shopify stores it in a machine-readable form.
Option / variant
A structured choice a product is sold in, such as colour or size. Options are the fields an agent can hard-filter on.
Filterable
An attribute exposed as a structured option, so an agent can narrow results to it (for example, filter to size 9). The strongest state, and the rarest.
Findable
An attribute written somewhere in the product text. The agent can read it and match on meaning, but cannot filter by it.
Missing
An attribute stated nowhere in the listing, so the agent cannot see it at all. The real gap.
Indexed
Whether a product exists in the catalog agents search at all. A product that is not indexed is invisible before readability even matters.
Inference (tech specs)
Specifications Shopify's own machine-learning model reads out of a listing's text. If even the inference is empty, no signal was there to read.
Precision
Of the top ten results an agent returns for a query, how many fit everything the shopper asked for. Higher is better.
Negation
A request that rules something out, such as without diamonds or not made of linen. Agents match on meaning rather than filtering, so they honour these only about half the time: a "not linen" search still returns linen.
Agentic checkout
An agent completing a purchase natively, without handing the shopper back to the store. Almost no products support it yet.
Every figure here was produced through Shopify's public catalog interface, the same one AI shopping agents read. The method is reproducible; a merchant can run their own catalog the same way at gpt.atomz.ai.
