Indeed Built an AI That Picks Better Job Matches for You

Job hunting on big sites often feels like shouting into a void: you apply to dozens of postings and hear nothing, while recruiters drown in hundreds of resumes that don't fit. Indeed's research team tackled one specific piece of that problem — deciding which candidates to show a recruiter first, and which jobs to show a job seeker first.

Their trick is surprisingly simple. Instead of asking an AI to write a paragraph explaining why someone fits, they ask it to pick a number from 1 to 5, like a five-star rating. That score takes one quick step of computing, so it's fast, consistent, and doesn't need any cleanup. They call it "single-token expected-value scoring" — think of it as reading the AI's gut reaction as a number rather than making it explain itself.

The hard part in hiring tech is usually the "cold start": new job postings and new candidates have no history, so big recommendation systems have nothing to learn from. Indeed's team showed a smaller, cheaper AI model can handle this if it's trained on a few hundred thousand human ratings of candidate-job fit. That's small by industry standards, but enough when the AI already knows general world knowledge about skills and roles.

In a live experiment on Indeed's own platform, employers saw 27.3% fewer irrelevant candidates and their "keep rate" — how often they moved someone forward — rose 7.07%. Simulated tests for job seekers showed even bigger gains. The catch: results come from Indeed's own platform and data, so the gains may not transfer everywhere, and a 1-to-5 score still flattens the messy nuance of whether a job is actually right for you.