AI Shoppers Buy From the Back of Your Catalog. Shopify's Own Numbers Say So

Shopify reported its third quarter on 3 November, and the headline figures were revenue and margin. The numbers that matter for anyone who owns a catalog came later in the call. Shopify's president, Harley Finkelstein, said AI traffic to merchant storefronts was up 2.6 times on a year ago and the orders from it up 3.3 times, and that 74% of AI-attributed purchases in the quarter came from outside Shopify's top 100 product categories. In the second quarter the figure was 75%. Jeff Hoffmeister, Shopify's CFO, put it as three quarters of AI-referred sales coming “yet again” from outside the top 100.

Shopify is careful to add that the volume is still small, and it is. But a share is a different kind of fact from a total. When three out of four purchases an assistant sends are for products outside the big categories, the assistant is not shopping the way your homepage does. It is shopping the way your back catalog has been waiting for someone to shop.

Specific questions find specific products

A field of hundreds of identical grey pebbles under cool light, with one oddly shaped stone rising on a narrow teal beam cut to exactly its outline. — A specific question has one good answer, and it is rarely the product that suits most people.

The reason is in how people talk to an assistant. Nobody types “jackets” into a chat. Finkelstein's examples on the second-quarter call were a car seat that fits three across a sedan, a reef-safe sunscreen that does not leave a white cast, a dog harness for a French bulldog and a standing desk for a small apartment. His phrase on the third-quarter call was that AI channels are “rewarding specificity”: they surface highly specific products that fit a specific customer with a specific need.

A specific question has one good answer, and it is rarely the bestseller. The bestseller is the product that suits most people. The product that suits a sedan with three car seats across the back is three shelves down, and it has been there all along. On a category page it is one thumbnail among forty. In a conversation it is the answer.

The traffic behaves accordingly. In the second quarter Shopify said half of all AI-referred sessions landed directly on a product page, two and a half times the rate of traditional search. In the third quarter it put the rate at three times organic search, with those shoppers converting 64% more often. An assistant does not send a shopper to browse. It sends them to the one product it has already decided on, and the shopper arrives most of the way to buying.

The product it decided on is the one you never polished

A single shelf of product boxes where the first few are richly painted and lit by a lamp, and the rest fade down the row into plain unpainted outlines in shadow. — Attention runs out long before the catalog does. The agents start where it ran out.

Here is the uncomfortable half. Every store polishes from the top. The twenty products on the homepage get the photography, the rewritten description, the size guide and the alt text. Product number 1,800 has the two lines the supplier sent in 2023, a title with the SKU in it and a photo called IMG_4471. The products an assistant is most likely to pick are exactly the ones that got the least attention, because they were never going to earn it on category traffic alone.

It matters more now because of what the assistant reads. Shopify's own explainer on Catalog, the structured product record it supplies to ChatGPT, Copilot and other agents, says Shopify “will infer missing attributes when the product data is clear enough”, naming colour, material, pattern, size type and age group. It asks merchants to use descriptions to answer real buying questions, with material, fit, dimensions, compatibility, care instructions, ingredients and certifications, and to put the most important facts early. “Fits three across” has to be in the record for the inference to find it. A description that says “great seat, ships fast” leaves nothing to infer from, and no theme design fixes that, because the theme is not what the agent reads.

An afternoon's audit of your own long tail

Sort your products by last year's revenue and skip the first hundred. That is the tail. Read ten descriptions at random from it and count how many answer a question a shopper would actually ask.
For each of those ten, write down the one specific fact a buyer would need before choosing it: the dimensions, the material, what it fits, who it is for. Check whether that fact appears anywhere in the title, the description or the options. Usually it is in your head and nowhere else.
Count the obvious gaps across the whole tail: descriptions under thirty words, descriptions shared between products, images with no alt text, products with no tags. Those are the fields an agent reads first and the ones most catalogs leave blank.
Fix the ten by hand, then look at the size of the list that remains. That number is the real job.

Finding the thin products without reading all of them

A floor of small glass tiles seen at an angle, a low lantern sweeping across them so the hollow ones flare teal as it passes while the full ones glow a steady amber. — Seven counts, ranked, and a run scoped to the products behind each one. The job is the list, not the reading.

This is where CortexCommerce fits, and it fits at the step most stores get stuck on: knowing which products in a long tail are thin. The CortexCommerce dashboard counts seven common catalog gaps across your live products: no description, a description under thirty words, a description shared with another product, no photo, images missing alt text, no meta description and no tags. It ranks them largest first, with a bar showing how much of your catalog each one touches. The counts are exact, not sampled, and they move as your catalog syncs.

Every gap has a button that opens an enrichment run already narrowed to just those products, so you are not re-drafting the twenty you polished by hand. Runs can also be scoped to a vendor, a product type, a tag or a collection, with a limit per run and an option to skip anything already enriched. CortexCommerce reads the product data and images you already have and drafts the missing fields in your brand voice: descriptions that say what the product is, who it is for and what it is made of, unique SEO titles and meta descriptions, consistent tags, and alt text written from what the AI actually sees in each photo. Every draft waits for your approval, and price, SKU, barcode and inventory are locked.

By hand, finding the thin products in a 3,000-product catalog means reading 3,000 descriptions. With a ranked list of gaps it means reading seven lines, then approving drafts for the products behind the biggest one.

We cannot promise that an assistant will pick your product; nobody honestly can. What Shopify's numbers do say is where the assistants are already shopping, and it is the part of your catalog you were least likely to fix. To see how your long tail reads today, run a free scan at cortexcommerce.ai/scan and start with the products it shows as thin.

Specific questions find specific products

The product it decided on is the one you never polished

An afternoon's audit of your own long tail

Finding the thin products without reading all of them

Sources