Users tell you what to fix. Their behavior tells you what to build.

At 9 users I had to learn to read feedback in two languages. What users say is one. What they do is the other. The gap between them is where most product work happens.

I launched a Chrome translation extension a few weeks ago. The first version optimized for accuracy. Better translations than the alternatives.

A user tried it and wrote me one thing: too slow.

I shipped a speed fix in the next few days. Translations stayed accurate, became faster. The specific complaint was addressed.

But that user's message taught me less than the broader behavior I was watching across the product. Words pointed to speed. Behavior pointed to workflow. Two different stories, same user base.

This is the most useful distinction I've picked up since launching.

What users say is one language. It reflects what they think they want, what they think you want to hear, what they can articulate in the moment. The signal comes through politeness filters and self-image filters and the limited vocabulary they have for describing their own workflow.

What users do is another language. It reflects where their actual time and attention go when nobody is watching. There are no filters. The behavior is silent and honest.

These two signals don't always agree. When they don't, the gap is usually where the product opportunity lives.

Two patterns I've started recognizing:

Words optimistic, behavior absent. The user wants to want your product. They like the idea. They don't fit the workflow. The fix is usually positioning, not features.

Words critical, behavior loyal. The user complains but uses it daily. The complaints are about edge cases. The core value is intact. The fix is usually polish.

The mistake I made early: I treated words as the primary signal because they were easier to collect. The user wrote a message, I had something concrete to act on.

But the message wasn't asking me to make it faster as fast as possible. It was saying, in the only vocabulary they had on hand, that the friction of the product was greater than the friction of their alternative.

The speed fix helped. It didn't change the deeper math. To get there I had to stop reading the message and start watching the session.

For early-stage founders specifically: at 9 users you can do this manually. You can pull up the session log for every user and trace what they actually clicked. At 90 users you need infrastructure. At 9 you just need the discipline to look.

Most of us don't. Words are louder.

Building Fenly.me - AI Translation Extension for International Teams and Remote Workers.