Why recruitment needs a better first filter
An experienced recruiter told us he would not use a reverse marketplace right now, because the market is already full of candidates. He is right. That is exactly why hiring needs a better first filter.
A recruiter turned us down with one line: the market is full of candidates. This post takes his side, then shows why a full market is exactly where keyword screening breaks, and what we measure instead, down to the match card a hiring manager sees.
A few days ago we asked an experienced recruiter to try Obvelum. His reply was honest, and it is worth quoting:
A reverse job marketplace is not something we would use. The market is full of candidates now.
It was a fair point, and one we kept coming back to. The more we thought about it, the more we realised it was the strongest argument for why Obvelum should exist.
Because he is right. The market is full of candidates. And a full market is exactly where the old way of filtering falls apart.
A full funnel is not a solved problem. It's a different one.
When candidates were scarce, the hard part of hiring was at the top: getting enough qualified people to apply. Most tooling was built for that world.
That world is gone. The hard part has moved. It is no longer about finding more candidates; it is about working out which of these hundreds deserve an hour of a human's time before that hour is spent. Volume was never the scarce resource. Filtered signal is. And the fuller the funnel gets, the more a trustworthy first filter is worth.
Why keyword matching stopped working
A keyword on a CV is the cheapest thing in hiring. Typing "Kubernetes" costs nothing, and it costs the person who has never run a cluster exactly as little as it costs the person who runs ten. A signal that is free to produce, and equally free whether or not it is true, carries almost no information. That was always the weakness of keyword-and-title matching: it rewards whoever writes the right words, whether or not the depth is there.
For years that weakness was survivable, because writing a tailored application still took effort, and effort filtered. AI auto-apply deleted that effort. One candidate can now fire off hundreds of keyword-perfect, role-tuned applications a day. A post that drew 80 applicants draws 800, most of them optimised to match the exact words in the description. So companies buy AI screeners to filter the AI applications, and the keyword arms race accelerates on both sides. The matching logic is not broken. The input has been debased to free tokens.
What we actually match on
A filter is only worth anything if it measures something a candidate cannot fake by editing a word list. So that is what we built.
We do not check whether a skill appears. We compare each skill the role requires against the candidate's proficiency level and years of experience. A skill that is present but below the level the role needs shows up as a partial match, in amber, never a green hit. You cannot keyword-stuff your way to coverage, because the depth has to clear the bar before it counts.
We separate must-have skills from optional ones. Piling on optional buzzwords does nothing for must-have coverage. Miss a high-level must-have and the score reflects it, however long the skill list is.
We roll individual skills up into semantic families, built from a map of 1,955 skills grouped into 71 families. Coverage is measured across a whole family, so one keyword borrowed from a domain does not earn credit for the domain. Some skills are inferred from where a candidate sits in that map rather than self-declared, and we mark those, because an inference is weaker evidence than a demonstrated skill and we would rather show that than hide it.
And the output is not a single score to game. It is a gap-and-strength breakdown: where the candidate's expertise is strongest, and exactly where it ends relative to what the role asks for. This is what a hiring manager sees against one of their own job posts:
Notice what the card refuses to do. It will not call a partial a match, and it will not let a long tail of optional skills paper over a missing must-have. Skills assessed by our system carry a distinct marker from skills the candidate simply declared.
Level, experience, breadth across a family: you cannot manufacture any of that in a word processor.
None of this has to be perfect to be useful. A self-declared level is a claim. A level that survives an assessment, sits next to years of experience, and lines up with the rest of a skill family is a much harder thing to fake. The bar was never perfection. It is whether comparing demonstrated depth produces a stronger signal than counting matching words.
The honest part: the pool is smaller
We will say this plainly, because pretending otherwise loses a recruiter's trust in one sentence. Obvelum is not where you go for more candidates. The pool is smaller. Everyone in it is anonymous until they choose to reveal themselves, and only visible if they have been active recently.
That is the trade, and we think it is the right one. A smaller set of people, each carrying signal that costs something real to produce, beats a flood of profiles tuned to the same job description. Ten candidates you can actually compare on depth are worth more than eight hundred you have to dig through to disqualify.
What this changes
For a recruiter, the first screen stops being keyword triage. You start from a shortlist ordered by genuine fit, you can see at a glance where each person is strong and where they fall short, and you spend your judgement on the conversation instead of on filtering noise. Reaching out costs credits, so every message you send is a deliberate decision, and every candidate who hears from you knows you were not blasting a list.
For a strong candidate, it means not being passed over because their CV said "server-side development" where the job said "backend." They are matched on what they can actually do, with more context than a keyword score can hold.
Reading the objection again
"The market is full of candidates" is not a reason to skip a better filter. It is the reason you need one. When discovery is solved and the funnel is overflowing, the work is no longer finding people. It is telling, quickly and reliably, which ones matter.
The market is full of candidates. What hiring teams are short on is a first filter they can trust. That is the part we built.
After this post went out, we dug into Stanford's audit of a shared AI screener: 4 million applications, one vendor, 150 employers. It takes the argument here one step further. A weak first filter does not just waste time; it rejects the same people everywhere at once.