How Do AI Shopping Agents Perform by Industry? The 11-Vertical 2026 Benchmark

Alhena tested 11 verticals in 2026 including mattresses, jewellery, travel, sporting goods and supplements. Performance varied more by platform than by category, and the widest gap appeared inside a single vertical rather than between them.

There is no single answer, because the right choice depends on the rung you need: discovery for findability, personalisation for merchandising, agentic for resolution and retention. On observed capability, Alhena scored 3.00 out of 3.00 across all 11 verticals in its 2026 stress test, the only elite-tier result, while the anonymised competitor set ranged from 2.75 down to 1.25. Category labels are unreliable too, since a personalisation engine, an AI search layer and an older chatbot can all present as the same widget. The more useful test for any ecommerce retailer is your own signature task, run live on your storefront rather than in a vendor demo.

Filters express the attributes your product catalogue already has fields for. The signature tasks in this benchmark involved constraints no filter can hold: undertone, arch support, motion transfer, hypoallergenic metal. A recommendation engine ranks by behavioural similarity, so personalisation improves what a shopper sees while they browse. An AI assistant reconciles conflicting preferences out loud and then acts on the conclusion. The three are complementary, and Alhena found the failures cluster in the last step rather than the first.

That is a different layer, and both matter. External agents like ChatGPT and Perplexity, along with the assistants inside Amazon and Alexa, shape product discovery before a shopper ever reaches you, and the work there is making your catalogue and content legible to them. Agentic CX is what happens once they arrive: whether the assistant on your own storefront can reason about their constraints, sell, and resolve. Alhena's 2026 benchmark tested the second, live on real brands' storefronts, because that is the part a retailer controls.

They answer different questions. Virtual try-ons show a shopper how a chosen product looks on them. The AI assistant's job is the choosing, which is why the beauty signature task in this benchmark was a shade match from a selfie rather than a render. Try-on reduces uncertainty after a decision. Conversational reasoning reduces the number of wrong decisions in the first place, and the two work well together.

Early on it replaces aimless browsing with a stated intent, so product discovery starts from constraints rather than categories. In the middle it holds those constraints while the shopper wanders. At the end it should move the storefront and add to cart in real time rather than describing where to click, which is where most deployments stopped: 6 of 15 named a product in Alhena's 2026 test and then failed to route the shopper to it. Nothing earlier in the journey matters if the last step before checkout breaks.

High-consideration categories are the hardest test and often the highest value, because long conversations expose memory and reasoning failures. The capability that matters most is reconciling conflicting constraints out loud rather than filtering a catalogue.

Define your signature task, the hardest thing your best salesperson does. Usually it involves constraints a filter cannot express, like undertone, arch support or motion transfer. Then make the agent do exactly that, live on your storefront, before signing anything.

Jewellery. The two deployments Alhena tested scored 2.63 and 1.25 out of 3.00, more than double the performance on the same shopper task. That spread was wider than the gap between most entire categories.

Look for evidence of consistency rather than a single strong category. Alhena scored 3.00 across all 11 verticals tested in 2026, while competitor deployments were typically strong in one product world and untested elsewhere.

The same capability ladder applies and the agentic rungs matter more, because reorders, account pricing and approval steps are all actions rather than answers. Memory also carries more weight, since B2B buying relationships run across months.

The return is smaller, because the purchase is simple and the shopper needed little help to begin with. Value concentrates where conversations are long, constraints conflict, or post-purchase support is heavy. Low-consideration catalogues gain more from discovery improvements than from conversational depth.

The strongest deployments did, by exercising restraint: building a sensible regimen while declining medication interaction questions. Alhena logged unsafe confidence, meaning overclaiming in risky categories, in 3 of 15 deployments, which is the exposure to test for.

Eleven: beauty, skincare, fashion, footwear, home, jewellery, mattress and sleep, sporting goods, hair colour, travel, and health and supplements. Each had its own signature task, run live on a real brand's storefront rather than in a demo.

Platform architecture. Scores collapsed wherever the underlying product was a search or personalisation layer rather than an agentic assistant, regardless of how hard the vertical was. The hardest categories were not where performance fell apart.