Pipeline
Adding Tool Names Reorders Your Sourcing List Without Narrowing It
Sourcing terms narrow a list only when some profiles lack them. A tool name nearly every profile in your results already carries fails that test: adding ChatGPT, Copilot or Gemini reorders the same people. The terms that still cut name an artifact somebody produced or an obligation somebody held. Search for eval set, model card, postmortem, incident review, acceptable use policy, escalation path. Then check whether the last twenty people you sourced actually showed one.
The takeThe instinct to add another term is the instinct to keep the search feeling like work, and a recruiter under requisition load has every reason to want that. But a search string is the cheapest thing in the process to change and the least examined, a bad pairing. Treat the string as a claim about where evidence lives, write the claim down, and let the next twenty profiles say whether it held. Most strings would not survive that, which is the argument for running it.
Where Olive fits
Open a role and see what the work shows
A sourcing search reads what a person wrote about themselves. Once you have a shortlist, Olive reads a session: a 40-to-60-minute occupational assignment done with an AI assistant, returned as six findings with the timestamped excerpt behind each one, and the candidate is granted the same report.
Assess your shortlistWhy Did Adding ChatGPT Stop Narrowing Anything?
Because a search term separates people only when some of them lack it. A tool name that has reached nearly every profile in the pool you search changes the order of the results and not the membership. The string feels like work, the result count drops a little, and the same people come back wearing a different rank.
What moved is the claim rate.
LinkedIn's Work Change Report says the rate at which its members add new skills to their profiles rose 140% between 2022 and the report's January 2025 publication, across a platform of more than a billion people 1. The figure counts profile edits.
Profile edits are evidence about what workers think signals value. They are not evidence that anyone learned anything, and LinkedIn's membership skews white-collar and English-speaking, so the number does not describe the whole labor force.
A profile line is a self-report, and self-reports about AI have been checked against a test.
In February 2025, 288 Taiwanese teachers took both a self-rated AI literacy questionnaire and a knowledge test built on the same four-part framework. Their self-ratings and their test scores matched weakly: correlations ran from 0.07 to 0.24 2.
None of those people were job candidates, and a weak correlation is not proof that anyone overrates themselves.
The finding is narrow and it is still enough: what a person claims and what a person demonstrates are two different measurements.
Candidates know which words a sourcing search reads. How recruiters search, and how to be in the results is advice a motivated person can act on in an afternoon. A term that is cheap to add is cheap for everyone to add.
The same saturation arrives one stage later in the loop, where everyone passes the AI question in the screen and the question may have become a password.
Which Words Can Only Be Written by Someone Who Was There?
Terms that cost something to write down. A tool name costs nothing: anyone who has opened a chat window can add it truthfully. A term naming an artifact somebody produced, or an obligation somebody had to hold, can only be claimed by a person who was there. The profiles carrying it are a much smaller set than the profiles carrying ChatGPT.
Three families of term do that work, and each one names something that had to exist.
- Something was built and measured. eval set, golden set, offline evaluation, regression suite, retrieval, grounding, model card.
- Something broke and somebody wrote it up. postmortem, incident review, rollback, prompt injection, red team, guardrail.
- Somebody was accountable for the output. acceptable use policy, review queue, human review, escalation path, sign-off, disclosure line.
The three families read as engineering, and the requisition on your desk may not be engineering.
Indeed Hiring Lab counts distinct US job titles where at least five postings in a quarter carry AI or a related term in the employer's own title. It found 264 in the first quarter of 2022 and 822 in the first quarter of 2026, with 63% of them outside tech occupations 3.
The count is of titles, one per title however many postings carry it, so it measures how far the vocabulary travelled.
So build the non-engineering version out of the work itself. A claims reviewer's obligation word is second review, a marketer's is disclosure, a lawyer's is citation check, a clinic biller's is the code somebody signed. Pull them from the task list in your own requisition rather than from a vendor's skills taxonomy.
Write the String as a Hypothesis, Not a Bar
A string is a guess about where evidence lives, so write the guess down before you run it. Name the two facts a profile has to show before you send outreach: one artifact they produced, and one decision they were accountable for. Then the search either finds those facts or it does not, and you can tell which.
A bar is a rule you apply to everybody. A hypothesis is a claim you can be wrong about cheaply, which is what you want from a string nobody has ever validated.
Write the two facts in the words you would accept from the person, not in the words a taxonomy uses.
For a support lead: one queue or macro they changed after reading what the bot got wrong, and one escalation rule they own. For an analyst: one model output they rejected, and what they checked it against. For a marketer: one piece they refused to ship, and the reason in a sentence.
Both facts have to be visible before outreach, or the search is not doing the job you think it is doing.
The two facts also give the first message something specific to ask for, and the outreach message is where your AI bar is first stated whether or not anybody wrote it down.
Check What Your Last Twenty Sourced Candidates Actually Had
Open the last twenty profiles you sourced and mark which of them showed the two facts. The exercise takes half an hour and settles an argument the search cannot settle on its own. If two of twenty carried an artifact, the string is selecting on something else and nobody has looked at what. If nineteen did, the bar you wrote down is already the bar you use.
Record the ones you passed over as carefully as the ones you contacted.
The people a string never surfaced are the ones it cost you, and nothing in the results shows them. You cannot compute a selection rate on people who never applied, and a sourcing search runs before anybody applies at all.
A twenty-profile audit is the small version of a check somebody has already run at scale, on a hiring phrase whose deletion was tracked afterwards.
The phrase was the degree line, and 11,332 roles at large US employers took it out.
Matched against the career histories of roughly 65 million workers by the Burning Glass Institute and Harvard Business School, those roles show the share of hires without a bachelor's degree rising about 3.5 percentage points 4.
Where the words came out, the edit did real work.
Those roles were few. Only 3.6% of roles at those large firms dropped the requirement at all. So the national net effect came to 0.14 percentage points: about 97,000 workers a year out of 77 million hires 4.
Neither figure survives being quoted alone. One overstates the reach of the edit, and the other understates what happened inside the roles that changed.
A search string is that kind of edit too. Rewriting it is free, and what costs something is the reading and the outreach behind it moving with the words.
Who a Sharper String Never Reaches
A narrower string finds people who wrote things down, which is a different population from people who did the work. Public evidence of AI judgment comes from employers that permit publishing, in roles where writing is part of the job, from people who maintain a profile at all. Everyone else does the same work and leaves no trace a search can reach.
Qualified people get lost this way, which is the cost sitting against every narrower string.
In a Harvard Business School and Accenture survey of 2,275 executives in the US, UK and Germany, fielded in January and February 2020, the filtering question went only to those whose organization uses a recruitment management system. Of those employers, 88% agreed that qualified high-skills candidates are vetted out because they do not match the exact criteria in the job description, and 94% agreed for middle-skills hires 5.
Employers agreeing about their own filters is not a measured exclusion rate.
The fieldwork predates generative tools entirely.
A narrowed list is still a stack of documents, and AI skill doesn't survive the resume stage whichever words found them. What a sharper string buys is a shorter reading list.
The string you started with had a tool name in it, and every profile in the results had the same word back.
Take the tool name out. Put in one artifact word, one incident word and one accountability word from the occupation on the requisition, and write down the two facts you will accept before you press search.
The list gets shorter. This time you can say what it dropped.
Common questions
Do Boolean search strings still work for sourcing AI skill?
Yes for locating artifacts, no for measuring skill. A string is good at finding pages where a particular word appears, so it can find the person who published a postmortem, wrote an acceptable use policy or maintained an evaluation set. It cannot tell you whether that person exercised judgment, because judgment leaves no keyword. Use the string to assemble a list worth reading, then read the evidence behind each hit yourself. If a term in your string appears on nearly every profile you would have surfaced anyway, it is decoration and can come out without changing the list.
Should I search for Claude or Copilot instead of ChatGPT?
Swapping one tool name for another buys a smaller list, not a better one. A stack name tells you which product somebody had access to at their last employer, which is mostly a fact about that employer's procurement. Knowing it matters when the role needs day-one fluency in a specific product and the ramp is short. It is not evidence of skill, and a candidate whose whole history is one assistant may carry the judgment across intact. Search on the stack only when you can say what breaks if the person has to learn a different interface in week one.
Do AI certifications narrow a sourcing list?
A certificate narrows a list, though what it tells you depends on what its issuer tested. It is a claim with an issuer behind it, so it is harder to add casually than a skill tag, and that alone makes it a sharper filter than a tool name. If the issuer tested a course and a quiz, that is not yet work anybody depended on. Treat a certificate as a reason to read further, never as the fact you were looking for, and check whether the person produced anything after the course ended. A course finished in 2024 with no artifact since is weaker than one changed workflow with a date on it.
What should I do when the narrowed search returns almost nobody?
Loosen one term at a time and watch which one was doing the excluding. A string with six requirements in it has usually stacked conditions nobody meant to stack, and the term that empties the list is often a vocabulary accident rather than a real bar: one employer says eval set, another says golden set, a third says regression suite. Try synonyms before you try lowering the standard. If the list stays empty across every phrasing, the evidence you asked for may not exist in public for this occupation, which is a fact about the role.
Can a sourcing tool's AI-skill filter do this for me?
A filter built on profile text reads the same self-reported words you would read, one layer further from view. It applies whatever definition its vendor encoded, which you did not write and usually cannot inspect, and it returns an ordered list that looks like a judgment. Ask the vendor which fields the inference reads and what the ordering is computed from. If the answer is skill tags and job titles, the tool is fast at the thing that stopped working. Keep the list it produces, then check by hand how many of the top twenty carry an artifact you can open.
References
- 1. Work Change Report: AI Is Coming to Work ✓ economicgraph.linkedin.com Supports the claim that the rate at which LinkedIn members add new skills to their profiles rose 140% since 2022, cited as self-reported profile editing rather than as skill acquired.
- 2. How to Assess AI Literacy: Misalignment Between Self-Reported and Objective-Based Measures ✓ arxiv.org Supports the claim that self-rated AI literacy and a knowledge test of the same construct correlate weakly, r = 0.07 to 0.24, among 288 Taiwanese teachers surveyed in February 2025.
- 3. AI Is No Longer Just a Tech Occupation Story: It's Spreading Across Job Titles in the US and Europe ✓ hiringlab.indeed.com Supports the count of distinct US AI-touched job titles rising from 264 in Q1 2022 to 822 in Q1 2026, with 63% of them outside tech occupations.
- 4. Skills-Based Hiring: The Long Road from Pronouncements to Practice ✓ static1.squarespace.com Supports both the 3.5-percentage-point change inside roles that dropped a degree requirement and the 0.14-point net effect, used to argue that editing requirement text is not by itself a hiring practice.
- 5. Hidden Workers: Untapped Talent ✓ hbs.edu Supports the complication paragraph: among the surveyed employers who said their organization uses a recruitment management system, 88% agreed qualified high-skills candidates are vetted out for not matching exact job-description criteria, rising to 94% for middle-skills hires. The employer survey was fielded January-February 2020 across 2,275 executives in the US, UK and Germany.
5 sources, numbered by first appearance. Every one was opened and checked against the claim it carries. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.