Columns2026/09/24

[DotDev 2026 Report Vol. 3] Shopify Catalog — A Product Search Engine Built for Agents, Not for Humans

I'm Reona, an architect at Flagship. In July 2026 I went to Shopify DotDev 2026 in Toronto, Canada. The first report covered the highlights of the event as a whole, and in the last one Rossella wrote about Shopify's change of direction for Liquid. This time I want to take one session from that event: the one on Shopify Catalog.

E-commerce, until now, has been built for humans

Add more products and, somehow, fewer of them get found. If you run an online store, I suspect that feeling is familiar. The more SKUs you carry, the less your on-site search hits. So you add filters, build curated collections, rebuild the homepage. The work of keeping the shop floor in order grows with the number of products you carry.

The Shopify Plus page, the top tier of Shopify's plans, lists more than 10,000 checkouts per minute, alongside billion-dollar brands handling over a million products and variants with ease. None of that foundation loses its value because the age of AI has arrived. But even with a million products in one store, the one actually visiting has always been a human.

The Shopify Plus page. On a black background, the heading Trusted by millions of global brands and partners sits above six cards: R&D innovation, Load-tested checkout, Unlimited SKUs, Built-in security, Detailed reporting, Partner ecosystem

The Shopify Plus page: "10K+ checkouts per minute" and "Billion-dollar brands easily host over one million products and variants." As of September 17, 2026.

In the session on Shopify Catalog, Shopify's Kate Ragotte (Senior Staff Product Manager, Agentic Commerce) put it this way. There is a limit to how much data a person can handle, so we lean on brands we already know and trust. But once AI can compute a signal for taste, a brand the buyer has never heard of can be put in front of them, as long as it fits.

Once AI is the one searching and buying, the ceiling we assumed comes off. And when the way people shop changes, so does where a store operator should put their attention. Sitting in that session, I came away with the sense that e-commerce is shifting underneath us.

There has never been a search engine built for products

Shopify describes Catalog as a search API that AI agents call directly when they look for products. It already holds structured data on billions of products from the millions of businesses running stores on Shopify — merchants, in Shopify's terms. Catalog took two years to build, and was first announced on May 21, 2025. It started out for a handful of partners, and in a little over a year it has grown into every merchant's products being listed in Catalog by store default.

Why they built it themselves

In the Day 2 main-stage fireside chat, Harley Finkelstein asked CEO Tobi Lütke why he was so fired up about Catalog. Tobi's answer was that really, really wonderful companies are going to be built on top of Catalog. The explanation went on. Until now, anyone trying to build product search had to borrow tools designed for information retrieval. But products, unlike what an information retrieval engine is built for, carry concepts that do not reduce to formulas. How stylish something is changes depending on who is asking, and by place too — whether you ask in Milan or in Ottawa. Computers have traditionally struggled with concepts like that, but AI models trained on everything people have ever written can now handle them as numbers and matrices. So Catalog is Shopify building a search engine for products from the first line of C++. On stage he went a step further: the people who actually built the search engines we use every day are a very small group who worked on indexing at Google, Yahoo and the like — and those people, for the first time, became interested in building one for products.

The five apps Shopify introduced at the Spring 2026 Editions were all built on Catalog and UCP: finding products from something in a video, packing for a trip from the destination, telling you what you need from the day's horoscope, searching by image, recommending board games by voice. The entrances into Catalog are not only the large ones like ChatGPT.

On Day 1, in the opening main-stage keynote, Shopify's Vanessa Lee (VP of Product) introduced a capability in development for classifying products by vague attributes such as formality, occasion and comfort (Fuzzy Attributes). Right now, the categories AI is good at searching are ones like electronics, supplements and watches, where attributes are standardized and product pages carry a lot of them. If Fuzzy Attributes can turn the nuance and taste of a product into numbers from the product image, and handle them as matrix computation, agents will be able to search across far more products and find the right one for the person on the other side of the chat.

What changed is the entrance

Searching for products across stores was already possible with the Shop app. But the Shop app assumes a person opening an app. What changed is the entrance. The one doing the searching switched from a person to an agent.

You disappear from the shortlist before you are ever chosen

From here I want to walk through how product search actually works in Catalog, as it was presented in the Day 2 session "AI in Commerce, Shopify Catalog, and AI Search Visibility" (Kate Ragotte / Kyle Risley).

The session showed a diagram of how an agent turns a request from a person into product searches. A question — "waterproof hiking jacket, under 500 dollars, ships to Canada, holds up in a backcountry downpour" — is first broken down by the LLM into a concept (rain jacket), constraints (under 500 dollars, ships to Canada) and preferences (size M, a brand they bought before, a taste for niche brands). The preferences are not in the question. They were added from past conversations between that user and the AI.

From there, several product searches run at once. The intent is rewritten several ways, such as "backcountry rain jacket" and "Gore-Tex rain shell," and fired in parallel. A web search runs at the same time, and the products recommended by the review sites it finds get turned into searches by product name and join the parallel run. The few hundred results that come back are then reordered by the LLM, which decides what to show the user.

Diagram of an agent breaking a single sentence from a buyer into several parallel product searches. The filter box on the right is labelled as applying to all of them

One question splits into several product searches, and the filters at the right apply to all of them. From "AI in Commerce, Shopify Catalog, and AI Search Visibility" (Kate Ragotte / Kyle Risley), Day 2.

One missing field, and you fall out of every branch

What I want to draw attention to here is that the filter conditions apply to every branch. A product with no gender attribute does not fall out of one search; it falls out of every branch running at once. The slide put it as one missing field can result in zero visibility.

On a search engine, a product with thin attribute data used to sink along a continuous scale called rank. An agent works by cutting on whether you match the conditions. One blank, and you vanish before you ever reach the scale. That is what is different.

Slide titled One missing field can result in zero visibility. On the left, the answer from the agent; on the right, web search results

The slide made this consequence its headline. On the left is the agent's answer, setting its own spec conditions: three-layer construction, a proven waterproof membrane. From "AI in Commerce, Shopify Catalog, and AI Search Visibility" (Kate Ragotte / Kyle Risley), Day 2.

How far Shopify can fill the gaps

They also talked about Shopify filling in missing attributes with machine learning. But it completes, structures and standardizes; it cannot invent something from nothing. Improvements like Fuzzy Attributes, which pull information out of images, may well cover more of these gaps automatically over time. Even so, nobody knows a product's attributes better than the merchant, and filling that data in correctly matters more now than it used to — that, I think, is the takeaway.

Catalog has other touches of its own. So that the same product does not appear over and over across stores, for example, Shopify bundles them together. The criteria for that bundling are described in detail on Shopify's engineering blog, if you are curious.

You can reach one of the three

The Catalog session also showed a diagram splitting the data behind AI search into three, by how far a merchant can move it: what you control directly, what you can only influence, and what you can only adapt to. The headline was "Merchants do not have direct control over all data," with a note in the top right saying that if social proof and authoritative sources could be controlled directly, their value would be diluted.

Slide titled Merchants do not have direct control over all data. Three columns, Product Matching, What Makes a Good Product and What Shoppers Care About, sit above a Control to Influence to Adapt axis

Left to right: what you control directly, what you can influence, and what you can only adapt to. From "AI in Commerce, Shopify Catalog, and AI Search Visibility" (Kate Ragotte / Kyle Risley), Day 2.

The layer you control directly

The first is the layer a merchant controls directly. What they put first were the three sources of product data: the product page itself, the product data in Shopify Catalog, and third-party product feeds. The requirement is that all three are complete, error free, and not contradicting each other.

Within product data, every product page in the admin has a category metafield mapped to Shopify's taxonomy. It arrived with the standard taxonomy that landed in every store in June 2024, and you use it by fitting your products into categories Shopify has defined. Shopify's machine learning models fill in inferred values from five things: title, description, images, options and tags. Put the other way round, a blank field means that information is written nowhere in your product data. If the material field is empty, chances are the material is not written anywhere in your description. All but tags, incidentally, are also the fields passed straight to agents.

The session also gave the other side of it. You talk to an AI agent, filter on Gore-Tex, land on the product page, and nowhere on the page does it say the material is Gore-Tex — at which point a person is not going to spend 500 dollars. The fields in the back are tools for staying in the running. What decides whether someone buys is what is written out front.

They also talked about how much return and refund policies matter. It is one of the things agents ask merchants most often, and not having one affects both discovery and conversion.

The layer you can only influence

The second is the layer you cannot reach directly but can still influence. A phrase that kept coming up in the Catalog session was "micro-consensus," introduced as an idea they talk about a lot inside Shopify. The view is that the small agreements formed in communities all over the internet are being quoted directly as the grounds for an agent's recommendation. In an Ahrefs study of roughly 75,000 brands, mentions on YouTube correlated most strongly with AI visibility (around 0.74). That is a wide gap next to branded search volume (around 0.35), and sheer page count had almost no correlation at all. What the session underlined was that today's search rankings remain a strong input into whether an agent finds you. Not replaced — the audience just got bigger. This route shows up in the jacket diagram from earlier: a review site's recommendation, found by web search, turned into a search by product name and joined the candidates. Being mentioned by a third party does not just push your rank up. It makes your product name the agent's own search term.

What a merchant can do in this layer is hand buyers a phrasing they can repeat in their own words. The example given was the skincare brand rhode, which turned "hydrating skincare" into "glazed donut skin."

The layer you can only adapt to

The third layer. In the jacket example, the buyer's size and preferred brand were not in their own question; they were added from past conversations. This personalization layer is one a merchant can neither see nor move, and it is an important input into how agents search. Two buyers asking in the same words will have different searches run for them. The assumption that "where do we rank" has a single answer breaks down here. What you are left with is deciding how to fill in the range you can move, knowing there is a layer you cannot.

How buying from AI works has not settled yet

Being in the index and being buyable inside an AI channel are not the same thing. What we have looked at so far is the former; the latter is built on a different mechanism. Catalog sits in the admin as one of Shopify's sales channels, and if you decide the exposure is not worth it you can turn catalog access off (it is on by default for eligible stores, and you choose which AI channels to sell through under Sales channels > Agentic in the admin).

The Agentic sales channel in the Shopify admin, showing the number of shoppers AI agents sent to the store in the last 30 days, a per-channel breakdown across ChatGPT, Meta, Microsoft Copilot and Shop, and the status of product sync and policy compliance

The Agentic channel in the admin. You can see which AI channels sent shoppers, and this is also where catalog access is turned off — a store-level switch, not a per-product one. Figures blurred.

The buying side, on the other hand, has not settled. In the Day 1 UCP session "UCP, Shopify, and Google" (Mani Fazeli / Ilya Grigorik), the attempts that came before were listed with dates. Buy on Google shut down in 2023, and Meta's Instagram and Facebook Shops in 2025. The most recent is OpenAI. It launched Instant Checkout in September 2025, completing payment inside ChatGPT, then folded it within six months and announced in March 2026 that merchants would use their own checkout while OpenAI focused on product discovery. The reason given was that the initial version of Instant Checkout did not offer the flexibility they were aiming for. Discovery on the AI side, buying on the merchant's side. Every one of them has landed there.

Shopify is building a standard called UCP (Universal Commerce Protocol) together with Google. The spec is public, and the participants include Amazon, Meta, Microsoft and Stripe. Amazon, the largest retailer in the world, has stepped onto a mat built by Shopify, who used to be the ones chasing.

On those earlier attempts, the diagnosis in the session was "not the wrong idea — they lacked the infrastructure," and they added that they had been involved in every one of them. What is building UCP is the side that has had twenty years of practice, with millions of merchants and trillions of dollars in payments. I rate its chances highly.

Diagram of four stages, Crawl, Walk, Run and Fly, in a row, with a red We are here marker between Walk and Run

"Journey to the moon." The road to an agent being able to buy, in four stages: crawl, walk, run, fly. We are on the boundary between walk and run — a cart can be assembled and handed to the merchant's checkout, but buying a specified order automatically is not there yet. From "UCP, Shopify, and Google" (Mani Fazeli / Ilya Grigorik), Day 1.

The order has changed

None of what you have built has stopped working. Search rankings remain a strong input. Kyle Risley, who leads SEO at Shopify, put it this way in the same session: if you have been doing SEO for a while, this should feel familiar — these are the same principles of good SEO, just applied to a wider cast of characters, some of whom behave a little differently than Googlebot and Bingbot. Vanessa Lee, having shown that 50% of AI sessions land on a product page, said the website is still in play even when AI is involved. How the site feels, how fast it is, the story it tells — that is what people read, below the surface, when they decide whether a brand can be trusted.

Does shopping really become agent-centred from here? I think it does. Typing in conditions, narrowing, comparing, adding to the cart, filling in a shipping address, filling in payment details — for the tedious half of shopping, handing it to an AI is simply faster. You can feel how well it fits the moment you use it. For now, though, buying via AI is still a sliver of online overall, and what was shown was slope rather than absolute volume: sessions up 8x year over year, orders up 13x. Those are Q1 2026 figures; in Q2 both settled to around 3x.

Of the three things Vanessa Lee put to the partners in the room at the close of the first report — Data, Taste and Distribution — what we have covered here is the first two. Writing out attributes and specs in full is data; holding your own way of saying things as a brand is taste. The third, the relationships and trust that only you have, I will leave for a later installment, but even if agents redistribute discovery, they cannot redistribute your relationship with the people who have already bought from your store.

People finish discovering inside what they already know. An agent can trust an unknown brand on its substance. From "AI in Commerce, Shopify Catalog, and AI Search Visibility" (Kate Ragotte / Kyle Risley), Day 2.

This article draws on Vanessa Lee's main-stage presentation on Day 1, the Tobi Lütke × Harley Finkelstein fireside chat on Day 2, and the Catalog session on Day 2. The first two are public as the keynote and the fireside chat, with automatic dubbing available. They are genuinely worth watching.