Alhena AI vs imbalanced-learn

Alhena AI sells shopping agents for ecommerce, and its feed is currently one piece of research being published a finding at a time. The 2026 stress test ran fifteen live AI shopping agents through real storefronts as ordinary shoppers: all fifteen could answer, nine could sell, four could complete a return or order change, and one remembered the shopper on a return visit. Four of the last five posts restate those same numbers from a different angle — the answer-to-act gap, memory, and now a taxonomy separating personalisation engines, AI search and agentic assistants by which ceiling each hits.

The taxonomy post is the tell: by naming three technologies that share a chat box and assigning each a hard ceiling — Recommend, Sell, Act and Remember — Alhena turns its benchmark into a category ladder with its own product at the top rung. Around that sit dated vertical censuses separating shipped assistants from announced intent, an attribution model, and comparison pages against AI visibility tools including Profound. None of this is product news; the last shipped features in the feed were the embeddable agents in July.

Expect the study to keep yielding one post per finding until it is exhausted, then a refreshed census or a second vertical on the same template; actual release notes will keep arriving only as launch posts between research runs.