Building a research voice for an AI company nobody had heard of

Every company in the category claimed to be an AI leader, and most had nothing to show beyond a product page. This one had genuine internal research on evaluating model reliability, and no habit of publishing it.

The difficulty was crowding: the volume of AI commentary meant that anything general would disappear.

We narrowed hard. Rather than commenting on artificial intelligence, the company would say useful things about one unglamorous problem: how you test whether a model is actually right.

Narrow subjects are easier to own, and reporters remember the person who is specific.

The engineering team published an evaluation methodology with the failure cases included, which is rarer than it should be. We took that to reporters covering the reliability beat, and to conference programmes rather than press lists.

Speaking slots did more work than pitches: two conference talks generated more inbound press interest than a quarter of outreach.

The company is now contacted directly when journalists need a technical view on model evaluation, which was the entire objective. Inbound enquiries replaced outbound pitching as the main source of coverage in the second half of the programme.