Out of 131 people LinkedIn's AI-Assisted Search handed me for a reliability engineer role, that's how many had the word "reliability" anywhere in their current title. Zero.
Let me walk you through exactly how I know that, because this is not a hot take, it is a real test, and one you can run yourself on any hard-to-fill role you have open right now, in about twenty minutes.
A real, currently open requisition landed on my desk: Reliability Engineer, at a national laboratory, electrical power systems, arc flash studies, predictive maintenance.
Straightforward right? Until you notice the title itself is a trap.
"Reliability Engineer" happens to also be what half of Silicon Valley calls a software job, a Site Reliability Engineer, otherwise known as an SRE. Nothing to do with nuclear facilities or electrical systems.
Same words, completely different professions.
I ran that exact requisition through LinkedIn Recruiter four different ways. In all four test cases I queried the same location, on the same day, so nothing but the search itself could explain whatever came back differently.
Typed exactly as the job posting says, word for word.
No judgment added, just the requisition's own corporate language.
30 results.
A real search string, built not from the posting's phrasing, but from how people who actually do this work describe themselves. The software they use. The specific studies they run. The certification that actually means something in this field. This is the Corpus Analysis methodology.
25 results, the tightest of the four.
Identical criteria, but this time typed directly into LinkedIn's AI-Assisted Search, the plain-English box the platform is actively pushing as the modern way to search. Describe the role like you're talking to a person, let the AI handle the rest.
131 results.
Identical practitioner language from search two, but instead of typing it as a string, entered it into LinkedIn's separate filter fields. Job Titles, Skills, point-and-click interface rather than typed logic.
703 results.
Search one and search two shared five candidates. Search two and search four shared another five.
That overlap is a big deal. Different construction methods, both hunting the same thing, kept landing on the same real people. That's convergence. And convergence is evidence those particular candidates are genuinely fit for the role, not an accident of search wording.
Search three shared zero candidates with any of the other three.
Not low overlap. Zilch. Peeling back results, the reason became obvious. Thirty-nine percent of the 131 worked at the national laboratory itself. Forty-two had "Electrical" somewhere in their title. None had "reliability" anywhere. The AI-Assisted Search took a plain-English description built around "electrical or facilities reliability engineer at an industrial plant, utility, or national laboratory," and it locked onto "national laboratory" and "electrical." It dropped "reliability," the one word that defines the job, and never told me. It just delivered 131 results as if the search had worked. NOPE!
That's no small miss. That's the tool nerfing the search without disclosing it.
Search four revealed a different problem. One that took a bit of excavation to root out.
At a glance, 703 seemed like an error, an absurd number for a highly specialized technical role in a not-densely-populated region. But there was no error. Three separate causes stacked atop each other, and all of them with identifiable evidence.
One job title contained the exact phrase "Reliability Engineer" but preceded by the name of an internal software system, at a company in an entirely unrelated industry, food manufacturing. No electrical work. The filter can't tell the difference between an "Electrical" Reliability Engineer and that result, because both satisfy the same simple rule: Reliability and Engineer, adjacently located somewhere within the title. Words before (or after) the pair never get evaluated.
Another current job title has nothing to do with reliability, in the least. Several employers back, this person held a title that did contain the exact phrase. LinkedIn's search, by default, doesn't just look at what someone has on their current title, it matches their entire job history. A title from years ago can still pull someone into today's results, with nothing flagging that's what happened. Yes, you can specify if you want to match current titles only, or past titles only, but the default is simply both.
A third false positive result matched on exactly one thing, a skill listed as Root Cause Analysis. Nothing about reliability anywhere on the profile, current or past. That's because Root Cause Analysis isn't specific to this field. A huge range of professions use that skill, including plenty of people who have never touched a power system. A skill term that generic can't actually distinguish one candidate from another. The tag isn't wrong. It's just meaningless as a way of distinguishing qualification for this particular job.
Going one step further, I restricted search four to current titles only. With that the count dropped from 703 to 157, more than three quarters of it gone from one setting. That confirms how much of the noise was old job history being matched. But 157 is still six times larger than search two's 25, built from the identical words. Even correcting for history, the point-and-click version stays far looser than a string you type yourself, because typing lets you require several things at once. Filling in separate boxes tends to ask for any one of them, even when that's not what you meant.
It isn't "don't trust LinkedIn."
LinkedIn built something real here. Their own engineering description names a system that blends several kinds of search at once, the keyword matching search has always used, a newer layer that tries to understand meaning instead of exact words, and a ranking system on top tuned to what gets a recruiter to actually click and message someone.
That's a legitimate thing to build.
But, the problem enters when you lean on the "easy button" version and stop doing the part that used to be your job.
Every piece of advice circulating right now on how to use LinkedIn is aimed at candidates, advising them how to rewrite their profile so this new system finds them. But who is telling talent professionals the other half of that story, that the convenient version of this shift can fail completely on a role with an ambiguous title, and an ambiguous title is closer to the rule than the exception?
Building your own search from real, specific language, and tools people use, the way they describe their own work, does work the plain-English box does not reliably do. It requires everything at once instead of settling for any one thing. This test is the proof of that, not an opinion about it.
You don't need my exact job posting to check this yourself. You need one real role you already have open, ideally one with a title that means more than one thing, and about twenty minutes.
Run it the way you normally would, whatever you'd type straight off the requisition. Then build one honest search from the language practitioners in that field actually use about themselves, not the posting's language. Run those same words again, but in the filter boxes instead of the keyword field, and watch how far apart the two counts land. If your seat has the AI-Assisted Search prompt, run that too, and actually read the first page of names it hands you, not just the count.
You'll learn more about how your own results get built in twenty minutes than any feature announcement will tell you.
I teach the fuller version of this, the vocabulary work underneath all four of these searches, inside the AI Sourcing Method at TSI University. You don't need the course to run this test. You need a real job and the willingness to check your own work instead of trusting the box.
