Your job search tool is measuring the wrong thing

Every product in this category publishes how many roles it showed you. None of them publish how many you could actually accept. That gap is the whole industry.

Ask any job platform how it is doing and you will get a number about volume. Roles in the index. Applications sent this month. Matches surfaced today. All of it real, all of it measurable, and none of it the thing a person out of work actually cares about.

The number that matters is smaller and much harder to face. Out of everything you were shown this week, how many could you have realistically accepted? Not found mildly interesting. Accepted. Had the right to work there, could physically be there, spoke the language the job is done in, would not be disqualified in the first thirty seconds by something stated plainly in the posting.

Nobody publishes that number. I understand why. It is embarrassing, it is expensive to move, and volume is so much easier to grow.

Volume won because volume demos well

A feed with four hundred roles in it looks like abundance. It photographs well in a pitch deck. It gives a new user something to scroll on day one, which is the only day most products are judged on.

Relevance has none of those advantages. It shows up as a shorter list, which looks like less product for your money until the moment you notice that every item on it is worth your afternoon.

So the industry optimised the number it could move. And the people using it learned to compensate by applying to everything, because if the tool cannot tell a real match from a plausible one, volume is the only strategy left. Which floods employers, which makes them lean harder on filters, which makes it harder for good candidates to get read. The whole loop is downstream of one measurement choice.

A feed you trust is worth more than a feed that is full. That sounds obvious right up until you try to build it.

Relevance is unglamorous engineering

We build Jobsome, which prepares job applications for people. So this is not a neutral observation for us. It is most of what we work on, and it has been consistently harder than the parts that look impressive in a demo.

Three things make it hard, and none of them are about artificial intelligence.

Hard constraints are not preferences. A job requiring fluent German when you do not speak German is not a weak match, it is not a match. Neither is an on-site role in a country where you have no right to work. Most systems fold these into a score, so they come back as a sixty-two percent instead of being removed. The correct behaviour is to drop the posting before scoring it at all, and to be conservative about it: act only when the posting states the requirement plainly, and leave anything vague alone. A filter that guesses is worse than no filter, because you stop trusting the ones it got right.

Your data is dirtier than you believe. Here is one we caught in our own system recently, and it is instructive because it is so stupid. We were classifying a posting as remote if the word remote appeared anywhere in it. That included "this role is not remote." It also included "provide remote desktop support in our Riyadh office." Now consider what remote means downstream: a remote job is location independent, so it skips location matching entirely. Every false positive became a job that could reach anyone, anywhere, forever. One lazy pattern match at ingestion, and the location logic three layers up was quietly meaningless.

You cannot say "only here" with a blocklist. Almost every tool lets you exclude places one at a time. Almost none let you simply name the two or three countries you actually work in. The distinction sounds pedantic and it is the difference between a filter that works and a chore that never ends, because the world has rather a lot of countries in it and you will not finish listing them.

The other half is trust, and it runs both ways

Filtering jobs for a person is only one side. The employer has the same problem pointed in the opposite direction, and it is worse, because everything in front of them is self-reported.

Consider the reference. Anyone can write you a glowing one. A recruiter reading that you were a pleasure to work with cannot tell whether it came from a former director or from your cousin, so they discount all of it and fall back on a phone call they will probably never make. The reference is not broken because of how it is written. It is broken because nothing behind it is checked.

The fix is not better wording, it is verification, and the discipline it demands is that you never blur the label. If a recommender was confirmed through their company email domain, say so. If they were not, say that instead, plainly, in the same place. The temptation to make an unverified claim look a little more solid than it is will kill the value of every genuine one on the platform. A trust signal is only worth something while it is still hard to get.

Audit your own relevance, out loud

All of this decays. Sources change their formats, a well-meaning default gets flipped, someone ships a pattern match at two in the morning. Relevance is not a feature you complete, it is a standard you hold, and the only way to hold it is to check your own work on a routine, rather than waiting for a complaint.

You will not get the complaint anyway. Nobody writes in to say "this is technically functioning and completely useless to me." They just stop opening it.

So we run a relevance audit as a routine. We take a real feed, and we go through it asking one question per posting: could this person actually take this job. Anything that fails gets traced back to the line of logic that let it through. That is how the remote classification above surfaced. Not from a bug report, from the routine.

Being small is the advantage here

I want to be honest about why we can do this, because it is not discipline and it is not talent. It is size.

At a large platform, relevance is not one person's job. Ingestion belongs to one team, matching to another, the feed UI to a third, and the metric each of them is measured on is a local one. Nobody is accountable for the question the user is actually asking, because that question spans all three teams and lands on none of them. A cross-team audit of end-to-end match quality needs planning, headcount and someone senior enough to force it. It gets scheduled quarterly, if it happens at all.

We can look at the whole path in an afternoon. Ingestion, classification, filtering, scoring, presentation, all of it, held in one head at one time. That is the one structural advantage a company our size has over a company a thousand times our size, and it is worth more than it sounds, because relevance failures live precisely in the seams between those stages. The remote bug was not in the matcher and it was not in the filter. It was in the handoff, where a field meant one thing on one side and something else on the other. Those are exactly the failures a large organisation is worst at seeing.

So we spend it. Not as a marketing line, as the actual routine.

The one change that has done the most for us is showing people what we removed. Not a count, the actual postings, each with the sentence that triggered the exclusion and a way to disagree. It is uncomfortable to build, because it publishes your mistakes to the one person guaranteed to notice them. It is also the fastest quality feedback we have ever had, and it turns a filter from something done to a user into something they can inspect.

The shorter list is the product. Everything else is inventory.

Ramzi Hobeishy is the founder of Silent Key Technologies, which builds Jobsome, BurnKey, Fontbench and TagTitan.

---

Also published on LinkedIn on 27 August 2026: Your job search tool is measuring the wrong thing