An AI candidate match score tells you how closely a profile overlaps with the request you typed. It does not tell you how well the person will do the job. Those are different questions, and most hiring tools let you confuse them by printing the first one as a precise-looking percentage.
When we built Sol, the assistant on our homepage, we made a deliberate choice: Sol shows a match band ("Strong", "Good" or "Partial"), never a number like 87%. This post explains why, and how to read any match score you run into.
Key Takeaways
- A match score measures fit to your words. Job performance is a separate thing it can't see.
- Percentages imply precision no hiring method has. Even structured interviews, the best predictor we know, sit around .42 validity.
- A score is only as honest as the profiles behind it. Unverified skills produce confident, wrong scores.
- Use scores to sort a long list into a short one. Then stop looking at them.
What an AI Candidate Match Score Actually Measures
Under the hood, most match scores add up a few things:
| Input | What it captures | What it misses |
|---|---|---|
| Skill overlap | Does the profile mention React, Node, Postgres? | Whether the person is actually good at them |
| Seniority signals | Titles, years, scope of past roles | Title inflation, small-company "senior" vs big-company "senior" |
| Availability | Can they start in your window? | Whether they'll want your project |
| Text similarity | How much the profile "sounds like" your description | Everything a resume can't say |
Notice what's missing from the left column: any measure of output. A match score knows what a person claims. It doesn't know what they ship.
That's why the pool matters so much. On our bench, an engineer only appears in a match after passing both an AI interview and a graded coding assessment. So when Sol says "Strong match," the skills behind it were tested rather than copied from a LinkedIn headline. A score computed over unvetted profiles is a score of how well people write about themselves.
Why a 94% Match Score Is False Precision
Here's the number that should make every hiring manager humble. The 2022 re-analysis by Sackett, Zhang, Berry and Lievens found that structured interviews were the single best predictor of job performance, with a validity of about .42. General cognitive ability came in around .31. Those are the best tools in the industrial psychology toolbox, and they explain a modest slice of who succeeds.
A match score built from resume text is a weaker signal than a structured interview. So what does "94%" mean? Mostly that 94% of the words lined up. Printing that as a percentage invites you to treat the gap between 94 and 89 as meaningful. It isn't. Both people might be great. Both might be wrong for you.
Bands are more honest. "Strong" says: this profile covers your required skills at the right seniority. "Partial" says: something important is missing, and here's roughly what. That's the level of confidence the data actually supports.
There's a fairness angle too. Resume text carries signals that have nothing to do with skill. In Bertrand and Mullainathan's field experiment, identical resumes with White-sounding names got 50 percent more callbacks than the same resumes with African American-sounding names. Any score built on resume text can inherit that. That's one reason Sol's pre-email cards leave out names and employers, and why a band with a plain-English reason beats a decimal point: it's easier to question.
A Concrete Version
A Series A startup is hiring one senior backend engineer: Python required, Postgres required, Kubernetes nice to have. A matching tool returns six profiles with scores of 96, 93, 91, 88, 84 and 79.
The tempting move is to interview the top three. Let's look at what the scores were made of:
- The 96 mentions Python, Postgres and Kubernetes 40 times across a long profile. Untested.
- The 93 has all three skills and a passed coding assessment, but is available in ten weeks.
- The 88 has Python and Postgres, no Kubernetes, passed a coding assessment, available now.
If Kubernetes is truly a nice-to-have, the 88 is probably your best candidate today, and the 96 is the riskiest. The score ranked them almost backwards because it rewarded keyword density and ignored verification and timing.
Rewrite the same list as bands with reasons and the picture fixes itself: the 88 becomes "Strong match: required skills verified, available now, no Kubernetes." That one line is more useful than the six numbers.
The Honest Counterpoint
Scores do earn their place at volume. If a public job post pulls 400 applicants, a score that sorts them into "look at these 30 first" saves real hours, even if it's noisy. SHRM's 2025 benchmarking puts average cost per hire at $5,475 for nonexecutive roles, so any honest hour saved at the top of the funnel is worth having.
The mistake is using a score past the point where it's useful. Sort with it, then drop it. Once you're down to a shortlist, the score has nothing left to tell you, and every minute you spend comparing 91 to 88 is a minute you didn't spend on a structured interview.
Frequently Asked Questions
What is a good AI candidate match score?
There's no universal threshold, because each vendor computes it differently. Ask what goes into the match score, whether the skills behind it were verified, and whether it separates required from preferred skills. A "good" score from an unvetted pool means less than a "Partial" from a tested one.
Can I trust an AI match score for engineers?
Trust it to sort, never to decide. Check that the skills were tested, then run your own structured interview. For a broader checklist on vetting, see hire pre-vetted developers.
Why does Sol show match bands instead of percentages?
Because the data doesn't support decimal-point precision. Sol shows "Strong," "Good" or "Partial" on blind profile cards, ranking the required and preferred skills and seniority you asked for against engineers who are vetted and currently available. Names and employers stay hidden until there's a signed agreement.
The Bottom Line
An AI candidate match score is a sorting tool wearing a lab coat. Read it as "how much this profile overlaps with what I typed," check that the overlap was verified, and do the judging yourself. If you want to see a band-based match on a real role, describe the role to Sol on our homepage. Meet Sol walks through what the cards show, and how we vet covers the tests behind every profile. For the mechanics of the matching itself, read AI talent matching.
Roberto Espinoza is CEO of Ruzora, which helps US startups hire pre-vetted senior LATAM engineers, with a vetted shortlist in 72 hours. See available engineers.
