The Filter You Can't See

90% of U.S. employers use AI screening tools to sort applicants. Most rely on the same small pool of third-party vendors.

You probably know about the keyword matching layer. You’ve adjusted your resume for it. You’ve pulled keywords from postings and translated your experience into the language the system expects.

There’s another layer most candidates don’t know about. And you can’t fix it with a better resume.

Retention prediction

Beyond keyword matching, many of these screening systems include retention models. They score candidates on estimated tenure probability, not just skill fit.

The model isn’t asking whether you can do the job. It’s asking whether you’ll stay long enough to be worth hiring.

These models infer age without accessing the field directly. Graduation years, total career length, and even older email providers serve as proxies. A model that predicts a high probability of retirement within five years for candidates past a certain career length can systematically downgrade those candidates for roles where the employer wants long tenure.

The candidate matches on skills. The model flags on tenure. A human never sees the resume.

The Stanford data

Stanford Digital Economy Lab published the August 2026 update to their “Canaries in the Coal Mine” report using ADP payroll data through June 2026.

The headline finding: employment of workers aged 22-25 in AI-exposed occupations is 19% below where it would be if it had kept pace with less-exposed peers.

Experienced workers show no comparable payroll gap. But the divergence operates through reduced hiring at the intake stage, not through increased separations. Employers are cutting at the front door.

For experienced candidates, the retention filter is the most plausible mechanism. Demand for experienced talent has been growing. Search timelines stay long. The disconnect lives at the screening layer, not in the market itself.

The compounding problems

71.4% of resumes score below the 75-point threshold most recruiters treat as the qualified cutoff. 82% of those rejections come from missing exact-match keywords, even when the candidate has matching experience.

The keyword problem and the retention problem are stacked. You can fix the keyword gap and still be scored down on tenure probability. Solving one doesn’t solve the other.

Then add the ghost job layer. Conservative 2026 estimates put non-actionable postings at 20-35% of listings in most sectors, with some analyses reaching 47%. The average ghost-job application cycle costs a candidate nine hours. For experienced candidates putting more time into each tailored application, the waste is proportionally higher.

Applications per hire increased 182% from 2021 to 2024. The average job seeker now submits 62.6 applications at roughly 44 minutes each, roughly 46 hours of application work alone. If a material share of those get rejected at the retention prediction layer before a human reads them, every hour spent tailoring the resume is spent on the wrong problem.

What you can actually change

You can’t see a retention prediction score. You can’t know which vendors a given employer uses. You can’t optimize your way out of a filter that runs before human review.

What you can change is the targeting strategy.

The retention filter is configured differently across employers. Some companies signal long-tenure preference explicitly: stable environments, slower growth, mature products. Others want senior execution for a specific phase and weight that over projected longevity. Internal referral paths and direct hiring manager contact frequently bypass the screening layer entirely, because the hiring decision happens before the ATS runs its pass.

The practical redirect: identify companies where the role signals match short-to-medium execution rather than 10-year tenure. Build the internal referral path before you apply. Make direct contact with the hiring manager before the application goes through the system.

Resume polishing is a keyword problem. This is a targeting and channel problem.

Getting clearer on which one you’re actually solving changes where you put the 46 hours.