The Most Important Signal in Shopify’s New AI Search Data Isn’t Traffic Growth
The practical takeaway from Shopify’s AI search data is straightforward: improve the product intelligence at the source, then make sure it can flow into the feeds, storefronts, marketplaces and other systems AI relies on to understand, compare and recommend products.
Shopify has published new data on how shoppers arriving from AI tools behave compared with shoppers arriving through organic search.
The headline figure is 197%. AI-referred sessions to Shopify storefronts grew 197% year over year in Q2 2026. Organic search sessions grew 12% from a much larger base, and orders attributed to AI referrals roughly tripled.
Those numbers will get most of the attention. But the more useful part of the report sits further down.
Shopify found that AI-referred shoppers performed especially well in research-heavy, specification-led categories, where people tend to compare compatibility, use case, technical details, reviews and trade-offs before buying. Across those categories, AI-referred shoppers converted at roughly twice the rate of shoppers referred through organic search.
Shopify also found that when AI search drew on structured Shopify Catalog product data, the shoppers it referred converted at around twice the rate of shoppers arriving from AI sessions that relied on scraped or third-party feeds.
That second finding needs some care. It is observational data. It does not prove that structured catalog data doubled conversion, and it would be a mistake to treat it as a simple cause-and-effect result. Merchant quality, product mix, and better source information appears to matter when AI is doing that work.
For ecommerce teams, that is a more useful starting point than other chart showing AI traffic going up.
AI is being asked to help with the choice
Traditional ecommerce search is well suited to direct product searches where the shopper already has a fairly clear idea of what they want. Take a search such as "men's black trail running shoes, size 10". The criteria are direct. Category, colour, size and gender do a lot of the work.
Now compare that with: "trail running shoe for narrow feet, 15-20km runs, wet and slippery terrain, and someone who tends to get heel blisters".
The second request asks much more of the system. It has to interpret several conditions at once, work out which product characteristics matter, compare possible options and assess whether the available information supports a recommendation.
A catalog can contain a complete set of standard fields and still be poor at answering that kind of request. Weight, heel drop, upper material and waterproof status are useful facts, but none of them alone tells you whether the shoe is a sensible match for that runner.
Shopify's data is interesting because the strongest AI conversion results show up in categories where shoppers tend to do more research before buying. That fits the way people are already using AI. They are not simply typing shorter product searches into a different box. They are handing over part of the research and comparison work.
Once that happens, the information available about the product has to carry more of the decision. A richer product record gives AI more explicit information to work from when that information is made available through the channels it uses.
Product records are usually strongest on facts
A typical ecommerce catalog might contain a title, SKU, price, colour, size, material, availability and a set of technical specifications. That data is necessary and shouldn't be viewed as outdated or secondary.
The weakness appears when a buyer wants to know whether the product suits a specific situation. Using the trail shoe example, the catalog might say:
Weight: 270g
Lug depth: 4mm
Waterproof membrane: No
Upper: Engineered mesh
Heel drop: 6mm
Those facts are useful, but the buyer may care more about whether the shoe holds well on slippery rock, whether the fit tends to run narrow, whether heel rubbing appears on longer runs, and whether the shoe is a poor choice for someone who needs waterproof protection.
At Canonica, we use product intelligence to describe that layer of the product record. It covers suitability, buyer fit, performance context, limitations, evidence, review signals and claim caution. The aim is to turn scattered product facts and evidence into structured information that helps explain when a product makes sense and where it does not.
Some of that information may already exist inside a business. It might sit in reviews, technical documents, support tickets, merchandising knowledge or supplier material. AI can draw on some of these sources when they are accessible, but scattered or vague information leaves more of the interpretation to the system. Turning important signals into clear product-level information makes them easier to use consistently across the channels where products are found and compared.
Creating the attribute is only part of the job. What you put inside it determines whether it is genuinely useful.
A structured field can still be weak
Shopify recommends capturing more detail around audience, use case, compatibility, fit, materials and “best for” scenarios, then structuring that information so it can be used by AI systems.
That direction makes sense. The risk is turning it into another field-completion exercise.
Imagine a catalog with values such as:
Best for: everyday use
Fit: comfortable
Terrain: outdoor
Use case: versatile
Nothing is technically broken. The data is present and machine-readable, yet it settles very little of a real buying decision.
The harder part is deciding which attributes matter for the category, how specific the values should be, what evidence supports them and where the brand should stop short of making a claim. There is judgement in that work.
For a trail running shoe, “good grip” may sound adequate on a product page. For a system trying to answer a question about wet rock, it is weak evidence. Grip can mean loose dirt, mud, dry rock, gravel or mixed terrain. A broad claim does not tell the system which condition it applies to.
Reviews can add useful context, but they introduce another problem. Ten positive comments about traction do not automatically support a wet-rock claim if none of the reviewers mention wet rock. A few clear comments about heel rubbing on longer runs may matter more than a much larger volume of generic comfort comments for someone prone to blisters.
This is the part that advice to “add more attributes” usually leaves out. Someone still has to decide what the attribute means, what evidence supports it and how far the claim can reasonably go.
Completeness can create false confidence
There is a familiar pattern in catalog projects. Teams find missing fields, populate them, improve schema coverage, clean feeds and move on. These are sensible tasks, and they often fix real problems.
They can also make a catalog look healthier than it is.
Imagine two trail shoes where both product records say "suitable for trail running", "good grip". "comfortable fit" and "durable construction". Every obvious field may be populated, yet the records still tell you very little about why one shoe should be chosen over the other.
This becomes more important when AI is doing the comparison. If both products are represented by the same broad claims, the system has little explicit information for deciding which one better matches a specific request. More complete records do not necessarily produce better recommendations if the information inside them does not distinguish one product from another in ways that matter to the buyer.
A more useful product record would explain that one shoe has stronger support for technical terrain and a narrower fit, while another is more forgiving for wider feet and better suited to mixed road-to-trail use. If evidence on wet rock performance is mixed, that should be represented as mixed. If reviews repeatedly mention heel rubbing after longer runs, that signal should not disappear simply because the official product copy says "all-day comfort".
The useful measure is how much uncertainty the information removes from a real buying decision.
Category knowledge matters here as well. Shopify's own data shows why. It notes that a T-shirt and a watch can sit within the same broad apparel category while requiring very different amounts and types of research before purchase.
The attributes that help someone choose a trail shoe will likewise be very different from those used to compare a camera, skin-care product or commercial coffee machine. Generic attribute frameworks eventually hit a limit because buyers don't evaluate every category using the same criteria.
Shopify's 2X finding needs to be read carefully
The structured-catalog result is probably the most commercially interesting part of Shopify’s report, and also the easiest one to overstate.
Shopify found that shoppers referred by AI converted at roughly twice the rate when AI search used structured Shopify Catalog data to find and recommend products, compared with AI sessions relying on scraped or third-party feeds.
A merchant cannot add structured attributes on Monday and assume conversion will double on Tuesday. Shopify’s data shows an association between the source of the product information and the performance of the traffic referred. It does not establish that structured data alone caused the difference.
The useful conclusion is narrower. The information AI has available when finding and comparing products appears to matter. That gives ecommerce teams a reason to look beyond whether their products are simply visible and examine what an AI system can actually determine about them.
A product can be indexed and available to an AI system while still providing weak information for a specific recommendation. If a shopper asks for a shoe suited to narrow feet and wet rock, knowing the SKU, colour, material and category does not resolve much of the request. The system needs information that connects the product to those requirements.
This is where I think a lot of AI visibility work stops too early. Visibility can tell you whether a product or brand appears. It does not tell you whether the available product information is good enough to explain why that product fits one buyer’s needs better than another.
Half of AI-referred sessions landed directly on product pages
Shopify also reported that 50% of AI-referred sessions in Q2 landed directly on product pages. It says these visitors are arriving deeper in the funnel and with more purchase intent than shoppers coming through traditional search routes.
Some of the research and comparison may therefore have happened before the shopper reaches the merchant site. A person can ask an AI tool for options, narrow the request, compare products and arrive at a PDP after already doing part of the work that once happened across search results, category pages and several product pages.
The PDP still has plenty to do. Price, stock, shipping, returns, reviews and the final product details can all affect whether that visitor buys. Shopify specifically recommends making PDPs useful for these deeper-funnel arrivals by including clear specifications, use cases, compatibility information, reviews and comparison-friendly detail.
There is also work happening before the visit. For an AI system to consider a product against a shopper’s requirements, useful product information has to be available in sources that system can read. Structured catalog data, feeds and product information exposed through the storefront can all contribute to that.
This is why I would be cautious about treating AI commerce mainly as a product-copy problem. Better PDP copy matters, but a business also needs to think about the structured product information it creates and how that information is passed into the channels where products are found and compared.
If the business knows that a shoe tends to suit narrower feet, but its structured product sources say nothing beyond fit = regular, that knowledge is difficult to carry consistently into feeds, marketplaces and AI shopping surfaces. The issue starts in the product record, even though fixing the record is only useful if the improved information is then made available to those systems.
A practical AI-readiness test for an ecommerce catalog
For years, teams have checked catalog health through field coverage, taxonomy, feed approval, title quality and specification accuracy. Keep doing that.
I would add another test. Take a sample of real questions buyers ask when choosing between products and check whether your product data contains enough specific, supported information to deal with them.
For a trail running category, those questions might involve wet-rock performance, foot width, longer-distance comfort, waterproofing, technical terrain or heel security. A camera, mattress or skin-care range will produce a completely different set of questions.
This matters because AI shopping queries tend to contain more of these needs and constraints. Shopify recommends looking at the language shoppers use across AI search queries, Search Console, ad search terms, on-site search and chat transcripts, then turning useful details into structured product data that AI systems can interpret.
Once you have the questions, trace where the answers currently live.
If answering a question about wet-rock performance means reading twenty reviews, checking a supplier PDF and interpreting a broad “excellent grip” claim on the PDP, the business probably knows more about the product than its structured record currently expresses.
An AI system can still encounter those sources. The problem is that it has to piece the answer together from scattered evidence. A clear product-level attribute gives the system something more explicit to work with and gives the merchant much more control over what is being stated about the product.
The exercise also exposes weak claims. If the product page says a shoe is “built for any terrain” while the available evidence only supports dry trails and loose ground, repeating all terrain = true in the structured record makes the problem worse. The claim has become easier for other systems to consume without becoming any better supported.
This is why evidence and confidence matter when building product intelligence. Suitability and performance attributes should reflect what the available evidence supports, including cases where the evidence is weak or conflicting.
For AI, the benefit is straightforward. When a shopper asks for a trail shoe for narrow feet, wet rock and longer runs, the system has clearer product-level information against which to evaluate each part of the request. It has less interpretation to do and better information for deciding whether the product belongs in the comparison.
A complete catalog doesn't guarantee that an AI system will recommend the product. It gives AI systems better raw material for making that decision.
What Shopify’s AI Search Data Means for Ecommerce Teams
The 197% increase in AI-referred traffic shows that AI commerce is growing quickly from a relatively small base. The more useful findings for ecommerce teams sit underneath that number.
AI-referred shoppers performed particularly well in categories where buying requires more research and comparison. Shopify also found that AI searches using structured Shopify Catalog data converted at roughly twice the rate of those relying on scraped or third-party feeds.
Those findings do not prove that better catalog data causes better conversion, but they give ecommerce teams a good reason to take the quality of their source product information seriously. Shopify is making the same point more explicitly in its current guidance for AI channels: important product facts should be structured and machine-readable rather than left for systems to piece together from whatever they can find.
I think the practical implication goes further than filling more fields. If AI is helping a buyer decide which product fits a specific set of needs, the useful information is often about fit, suitability, compatibility, performance conditions, limitations and evidence. Those signals need to be specific enough to help with the decision and well supported enough to trust.
The work starts at the source, where that product intelligence can be created and governed. It then needs to flow into the storefront, feeds, marketplaces and other channels where products are represented and AI systems can use it.
That is the part of Shopify’s report I would pay most attention to. The growth in AI traffic tells us the channel is becoming commercially relevant. The harder question for ecommerce teams is whether the information they are supplying is good enough for AI to do something useful with it.
Sources
Shopify Enterprise — AI-referred shoppers convert better and spend more: What Shopify’s early data shows
https://www.shopify.com/enterprise/blog/ai-search-category-behavior
Shopify Enterprise — 8 Tips to Prepare Your Product Data for AI Channels
https://www.shopify.com/enterprise/blog/product-data-tips-ai-channels
Shopify — What Is Shopify Catalog and How Does It Work?
Comments