January 7, 2026 · 5 min read
An AI Feature We Shipped That Nobody Used, a Postmortem
We spent six weeks building an AI-powered recommendation widget that got used by less than 2% of visitors, and the postmortem traced the failure to a decision made before a line of code was written.
The feature worked exactly as designed. An AI-powered 'find similar products' widget, trained on the client's catalog, returned genuinely relevant results in testing. We shipped it on time, on budget, with clean telemetry. Three months later, usage was under 2% of product-page visitors. This is the postmortem of why.
What we checked first
We assumed a visibility problem — maybe the widget was below the fold, maybe the copy was unclear. Heatmaps showed otherwise: visitors saw it, a meaningful fraction hovered over it, and then didn't click. The problem wasn't discovery. It was that the widget was answering a question nobody was asking at that point in the journey.
We built a good answer to a question our users weren't asking at that moment, and no amount of model quality fixes that mismatch.
The actual root cause
The widget was placed on the product detail page, right where a user has already decided to look closely at one specific item. 'Similar products' is a discovery-stage need, not a decision-stage need — at that point in the funnel, users want reassurance about the item in front of them, not alternatives to it. We'd built the right feature for the wrong moment in the journey, and no amount of recommendation-quality tuning was going to fix a placement problem.
- We never validated journey-stage fit before building — we validated recommendation relevance, which was never the thing in question.
- Usage data from a similar widget elsewhere on the site existed and we didn't look at it before scoping.
- The feature's AI component worked fine the entire time; the postmortem initially wasted a week investigating model quality before the team looked at placement.
We moved the same widget to the category browsing page, where 'similar products' matches what a visitor is actually trying to do, and usage went from under 2% to around 18% with zero changes to the underlying model. The lesson wasn't about AI at all — it was the oldest lesson in product work, just relearned with an AI feature as the expensive teacher.