A customer asked me last month how she found a kettle I sell. She did not type the product name into Google, did not scroll a category page, did not click an ad. She asked ChatGPT for a quiet pour-over kettle under $90 with a gooseneck spout, it suggested three, mine was one of them, and she clicked straight through to the product page. That is the whole funnel now for a growing slice of shoppers, and it has almost nothing to do with the SEO playbook most of us spent a decade learning.
I run a mid-size Shopify store and I have spent the last few months pulling my catalog apart to figure out why some products show up when I ask an AI assistant about them and others never do. The pattern is clearer than I expected, and it comes down to one thing: the models are reading structured product data, not prose. If your attributes are thin or missing, you are invisible to them, no matter how good your copy reads to a human.
AI search reads attributes, not adjectives
Classic SEO rewarded a well-written page: a keyword in the H1, a chunk of descriptive text, some internal links, backlinks pointing in. An AI shopping assistant works differently. When someone asks for a waterproof hiking boot in a wide fit under $150, the model is not skimming your paragraph about adventure and the great outdoors. It is trying to answer four discrete questions:
- Is this a hiking boot?
- Is it waterproof?
- Does it come in a wide fit?
- Is it under $150?
Those are attribute lookups. If your product has material: full-grain leather, waterproof: yes, width: wide, price: 139, the model can confidently include you. If that information lives only in a sentence buried in the description, or worse, is not stated anywhere, the model either skips you or guesses, and it tends not to guess in your favor. It would rather recommend a competitor whose data is unambiguous than risk a wrong answer.
This is the mental shift. You are no longer writing to persuade a reader. You are supplying facts to a machine that will decide, in a fraction of a second, whether you are a valid answer to a question you never saw.
The attributes that actually move the needle
Not every field matters equally. After tagging thousands of SKUs and watching which ones surfaced, here is where I would spend your time first:
- Core category and product type. Get this exactly right and specific. Not apparel, but mens merino base layer. The product type is the first filter the model applies.
- Material and composition. Shoppers constantly qualify by this (cotton, stainless steel, recycled, vegan leather) and so do the assistants.
- Dimensions and fit. Size, weight, capacity, width, length. These are the attributes that turn a maybe into a match.
- Compatibility. If your product works with a specific device, plug type, bike mount, or phone model, say so as structured data. This is huge for accessories.
- Use case and occasion. Waterproof for rain, dishwasher safe, carry-on compliant, machine washable. These map directly to how people phrase requests.
- Care, warranty, and certifications. Organic certified, BPA-free, two-year warranty. They build the confidence a model needs to recommend.
The common thread is specificity. Rich vague is less useful than modest specific. One accurate heel_height: 40mm beats a paragraph gushing about elegant silhouettes.
Why thin catalogs go invisible
Here is the uncomfortable part. Most Shopify catalogs are thin, and the owners do not know it, because the store still looks complete to a human visitor. A product with a title, a photo, a price, and two sentences of copy reads fine on the page. But strip it down to its structured attributes and there might be almost nothing there: a product type Shopify auto-assigned, maybe a color, and a vendor name.
To a human, that product is browsable. To an AI assistant, it is a near-blank card. When the model compares your near-blank card against a competitor who has fifteen clean attributes, you lose every head-to-head you are not explicitly disqualified from, simply because you gave the model nothing to say yes to.
I see three recurring gaps when I audit a store:
- Attributes trapped in prose. The fit, the material, the compatibility are all mentioned in the description, so the owner assumes they are covered. They are not, in any form a model can reliably parse at speed.
- Inconsistent values. One product says wide, another says W, another says wide fit. A human reads them as the same. A filter and a model do not.
- Empty metafields. Shopify gives you metafields and the standard product taxonomy precisely so you can store structured attributes. Most stores barely touch them.
How to fill the gaps without losing a month
The honest problem is volume. Enriching attributes by hand across even 500 SKUs is a soul-crushing week of spreadsheet work, and it goes stale the moment you add new products. That is exactly the job I now hand to our Scurix Product Attributes AI app, which reads each product and generates the structured attributes, metafields, and taxonomy data automatically, then keeps them consistent across the catalog. It turned a job I kept deferring into something that runs in the background, which is the only reason it actually got done.
Whether you automate it or grind through manually, the sequence is the same:
- Audit what you have. Export your products and look at the structured fields, not the pretty page. Count how many have a specific product type, material, and at least five meaningful attributes. The number will probably surprise you.
- Standardize your vocabulary. Pick one value for each concept (wide, not W or wide fit) and enforce it. This matters as much as filling blanks.
- Map to Shopify's standard taxonomy. Shopify publishes a product taxonomy with category-specific attributes. Aligning to it is the closest thing to speaking the native language that shopping surfaces and assistants expect.
- Fill metafields deliberately. Use them for the attributes your buyers qualify on. Think about the questions a shopper would ask an assistant, and make sure each one has a field that answers it.
- Add structured data markup. Product schema on the page (price, availability, brand, GTIN, aggregate rating) is still read by crawlers that feed AI results. It is not either-or with metafields; it is both.
Treat it as ongoing, not a one-off
The last thing I will say is that this is not a project you finish. Every new product you add starts thin by default. Every time an assistant gets smarter about parsing attributes, the stores with clean data pull further ahead. With BFCM traffic about to spike, the products that win the AI-assisted question will be the ones whose data was already in shape before the rush, not the ones scrambling to patch it in December.
The stores that treat product attributes as a living asset, as real as inventory or pricing, are the ones that stay findable as search keeps shifting under us. Thin catalogs were always a quiet liability. AI search just made the bill come due.
