Your ATS Archive Can Be Searched, Just Not for AI Words

Why does a keyword search of the archive return the wrong forty people?

Because the search matches words, and the words changed while the records sat. Indeed Hiring Lab put AI-related US job postings at 6.3% of postings in August 2026, well past a prior peak of 3.3% in 2022 ¹. Employers roughly doubled their own use of the vocabulary in four years. A candidate writing in 2023 was writing after that earlier peak and well before today's share.

That series counts employer language in postings, not resumes. No public series tracks how often applicants wrote about AI in any given year, which is the measurement this question needs and nobody holds.

Job titles drifted on the same schedule, and every file in the archive is stamped with one year of that drift. Indeed Hiring Lab counted distinct US job titles carrying AI or a related term in the employer's raw job title: 264 in 2022, 159 in 2023, and 822 by the first quarter of 2026 ².

A file dated 2023 was written in the leanest of those three readings for employer titles ².

Profiles moved too, and not necessarily because capability did. LinkedIn reports that the rate at which its members add new skills to their profiles has risen 140% since 2022 ³. Adding a fashionable skill to a public profile has obvious career value, so the figure measures what workers believe signals value rather than what they learned.

A keyword search finds the people who did that.

The mechanics make it worse than a rough proxy. Greenhouse's talent filtering searches the full text of resumes and internal notes, and states that a search keyword must exactly match the keyword in the application, with required keywords behaving as AND statements ⁴.

An exact string is not a concept, so Copilot, GPT-4 and prompt engineering are three separate terms over one person.

Greenhouse's rule is Greenhouse's. Systems differ on whether they stem a word, expand a synonym or match on meaning, so read your own vendor's documentation before trusting a null result.

On live applications the same filter fails in the opposite direction, which is the question behind what replaces an ATS keyword filter when every resume matches. The archive fails the other way. Almost nothing matches, and the misses are dated rather than unqualified.

What should the archive search actually filter on?

Filter on occupation, on recency, and on the specific tasks the open role needs done. Those three sit in old applications in a form that does not expire: what somebody did, where, and when the file was written. Rank nothing. The output is a list of people worth a message, and it deserves its own name in the ATS so nobody mistakes it for a shortlist.

Occupation is the field that survives. A person hired as a claims examiner in 2022 is still somebody who examined claims, and the work the open role needs done is described in the same nouns it was described in then.

Tool language does not survive, because the tool is what somebody happened to have.

Reconciled the intercompany accounts means in 2026 what it meant in 2023. Used ChatGPT meant one model in one text box in 2023, and means almost nothing on its own now.

Recency goes into the query as a date and stays a date. A file from 2022 is not weaker evidence about a person; it is less evidence, four years old, about somebody who has had four more years.

Search terms are their own problem, and the live-sourcing version of it is what terms actually narrow a sourcing list for AI skill. The archive adds one constraint: every term has to work on documents written before the term was common.

Search for the job, because the words will not be there.

Where does the AI question belong, if not in the search?

In the outreach, where the person can answer it now. A stored resume is a snapshot of a document somebody wrote for a different job on a fixed date, and nothing in it can be updated by reading it harder. One message asking what they have built with AI since applying produces evidence the archive cannot hold.

State the bar in the message rather than testing for it silently: the outreach message is where your AI bar is first stated.

Three questions are worth putting to a returning applicant, and none of them asks for a rating:

• What has changed in how you do this work since you applied.
• One thing a model produced that you threw away, and what was wrong with it.
• One claim from a model you checked against something outside the conversation.

Do not ask for a number out of five. A score somebody assigns to their own AI skill tracks how they feel about it, which is the problem with asking candidates to rate their AI skill one to five.

The answers come back current, which the file is not. They are also self-reported, which the file was too, so somewhere later a person still has to do the work in front of somebody.

When does a saved search over old applicants become a selection procedure?

The moment its output decides who gets contacted and who does not. The Uniform Guidelines define a selection procedure as any measure, combination of measures, or procedure used as a basis for any employment decision, and say the definition runs from paper-and-pencil tests through informal or casual interviews and unscored application forms ⁵. A saved query over forty thousand records plausibly sits inside that sentence, though the Guidelines date from 1978 and name no software.

The Guidelines attach a validation burden only where adverse impact appears, so the definition sets scope rather than a duty to audit every query somebody runs ⁵.

California has written a close case down. Its Civil Rights Council amended the FEHA employment regulations to cover automated-decision systems, effective October 1, 2025. The text names screening resumes for particular terms or patterns among its examples of a computational process that facilitates a human employment decision ⁶.

The regulations create no audit duty and no duty to tell anyone a search ran. An agent operating the system, a vendor included, is itself an employer under that Act ⁶.

Whether your own saved query lands inside those words is a question for employment counsel, and none of this is legal advice.

The record is the cheap part and the part nobody keeps. Store the query text, the date it ran, the fields it filtered on and the count of records it excluded, next to the requisition. A search nobody wrote down cannot be explained a year later, and the explanation is the thing people ask for.

Volume is the objection that stands. One recruiter with one open role and forty thousand records has to cut the list somehow, and a crude filter that runs beats a careful one that never gets built.

Nothing here removes the cut. What changes is the field the cut is made on.

Read the old interview notes as dated, not as wrong

Old assessment evidence in those records is dated rather than false, and the two call for different handling. A 2023 note saying a candidate was slow to check a model's output described a real observation with the tools of that year, and both the tools and what counts as competent use of them have moved since. Treat the note as a question to re-ask, never as a verdict to reinstate.

The rehire side of the same problem is written up in whether 2024 interview evidence still holds when you rehire from the pool. The archive version is older, and the drift is larger.

A rejection reason ages worse than an observation. Not enough Python is a fact about a requirement somebody wrote in 2023. Did not check where the number came from is a fact about how one person worked on one afternoon.

Retire the first kind, and re-ask the second.

The archive is worth what it was always worth, which is forty thousand people who once wanted to work here. Each record carries the month it was written and the job it was written for, and neither field has decayed at all.

Search on those, write to the people they return, and let the AI question be answered in 2026 by somebody who is around to answer it.

See a sample report >