Pipeline
The Parts of a Posting That Actually Reduce Generated Applications
Changing the posting can cut the volume of AI-written applications, but not the way the standard advice says. Adding requirements cannot deter a generated application, because requirements are what the generator writes from. Three things change who applies: a published pay range, an exact location and travel rule, and one short ask that cannot be mass-produced across a hundred postings. Allow AI on that ask and say so. Then watch applications a person actually read rather than applications received.
The takeThe volume panic is mostly a measurement problem in recruiting clothes. Applications received was always a proxy for candidate effort, and that proxy died the month drafting one became free. Teams answering it by adding requirements are bidding against a machine that writes faster than they can raise the bar. The posting still matters, but only the parts of it that are facts: money, place, manager, first project. The rest is decoration, and decoration is exactly what a model mirrors back.
Where Olive fits
Open a role and see what the work shows
Olive is priced per attempt instead of per seat, and one attempt returns six evidenced findings on a single candidate, written by a person: an input to your decision rather than a cut in your funnel. Ten attempts a month are free, so a short piece of real work can run beside the posting change and be compared against it.
Rank your shortlistWhy Do More Requirements Make the Pile Bigger?
Because a requirement is an input to the thing writing the applications. A candidate pastes the posting into a model, and every line you add comes back as a matching line in the resume. More requirements raise the apparent qualification of the pile without changing its size. Length behaves the same way: a longer posting produces longer applications, and longer applications cost more to read.
The second reflex, buying something that sorts the pile, has the same flaw one step later. A sorter changes the order in which volume arrives on your desk. It does not reduce how many people applied, and it decides on the one thing that stopped carrying information: how the document reads.
There is also no dependable way to sort applications by whether a model wrote them. Seven widely used detectors run over 91 human-written TOEFL essays produced an average false-positive rate of 61.3%, and 97.8% of those essays were flagged by at least one tool 1. Those were academic essays rather than resumes, and the tools have moved on since, but the direction of the error is the durable part: it lands on people writing in a second language. A screen built on that is not a smaller pile. It is a differently wrong one, and it is the reason keyword filters stopped separating anyone.
What is left is the front of the funnel. You cannot make writing an application expensive again. You can decide which candidate finds your posting worth fifteen minutes. That is a change in targeting, and it is the only lever the front of the funnel still has.
Publish the Range, the Location Rule, and the First Project
These are facts, and facts are the part of a posting a candidate cannot resolve by rewriting. A range removes the people the money does not work for. An exact location, days on site and travel share removes the people the commute does not work for. Naming the manager and the first ninety days removes the people who wanted a different job with the same title.
Ranges are not optional everywhere. California requires an employer with 15 or more employees to include the pay scale in any job posting, and to hand it to third parties posting on its behalf. That is Labor Code 432.3, most recently amended effective January 1, 2026 2. Other states and cities have their own rule with its own threshold and their own answer for remote roles, so check the jurisdictions where the work and the candidate sit, and check them with counsel.
Precision does the work, not disclosure on its own. A band running from $90,000 to $180,000 tells a candidate nothing and reads as a refusal. Hybrid with no number of days attached is the same. Write the version you would say out loud on a phone screen:
- Pay. The band you would actually offer at this level, and what moves someone inside it.
- Place. City, days on site, whether that is negotiable, and travel as a share of a month.
- Person. Who the hire reports to, how large the team is, who else is on it.
- Project. The first thing they will own, named down to the deliverable.
The last one costs the hiring manager ten minutes and is the line most candidates cannot get anywhere else. It also produces a posting a model cannot pad, because the specifics were never public.
Ask for One Thing That Cannot Be Mass-Produced
Add a single ask that belongs to this role and takes about fifteen honest minutes: a short answer to a real constraint the team is under, a read of a two-paragraph scenario, one artifact with a decision inside it. It does not work by refusing AI. It works because a candidate sending a hundred applications will not spend fifteen minutes on yours, and the one who wants this job will.
Say plainly that AI is allowed on it. What you are asking for is attention, and a model cannot spend that for anyone. An ask framed as a trap invites someone to beat the trap, and it teaches the candidate that the process is adversarial before anyone has spoken.
Keep it small enough to defend. Fifteen minutes unpaid is a reasonable request. Two hours is a take-home, and that belongs after a conversation with payment attached. If the answers are not interesting to read, rewrite the ask before you blame the candidates.
A job-shaped exercise also has research behind it, though weaker research than the usual citation implies. The 2022 re-analysis of the selection literature puts work samples at .33 and structured interviews at .42, revising down the .54 that still gets quoted for work samples 3. The credibility intervals behind those two numbers overlap heavily, so the honest reading is that a short piece of real work and a well-run structured interview sit in the same class of evidence. Almost every study behind those numbers tested people already doing the job, so treat it as support for the shape of the ask and nothing more precise than that. What to ask for instead of a cover letter works through the form of the ask in detail.
Change the Number You Watch
Applications received stopped measuring anything you control, so stop reporting it. Two counts replace it: applications a person read end to end, and how many of those reached a first conversation. A posting drawing four hundred with twelve real reads is not out-performing one drawing sixty with thirty. The second is a pipeline. The first is a queue with a dashboard on it.
Instrument it the cheap way. For two weeks, log what your team actually opened, what survived the first thirty seconds, and what got a reply. Cut it by source. If one board is contributing most of the volume and almost none of the reads, that is a spend decision you can make on Monday morning.
Two more things are worth watching once the posting changes:
- Completion of the ask. Does a smaller and better-fitting group finish it, or does everyone skip it? If almost nobody finishes it, the ask is too big or too vague.
- Questions about the band. Postings with a real range draw more questions about the range. That is the range doing its job.
None of this moves the count on the dashboard in week one. The reading hours move first, and those were the resource actually short. How long to leave the posting open is the other half of the same decision: at some point the answer is to close it and hire from what you have.
Common questions
Does writing 'no AI-generated applications' in the posting help?
It removes nothing and commits you to a rule you cannot enforce. There is no dependable way to tell which application a model wrote, and the tools that claim to do it misfire hardest on people writing in a second language. A candidate reading that line either ignores it or spends an hour making the output look handmade, which selects for the wrong thing. If AI use matters for the role, ask about it as work rather than as a prohibition.
Will publishing a pay range bring in more applications rather than fewer?
A published range usually brings in more of the right applications and noticeably fewer of the wrong ones. A range does not deter people who fit it, and it removes candidates the money was never going to work for, who otherwise reach a phone screen before anyone finds out. The volume effect is smaller than the effect on what the volume is made of. California requires it of employers with 15 or more employees under Labor Code 432.3, in force in its current form since January 1, 2026, and other states have their own rule, so the choice is often about precision rather than disclosure.
Where should the extra ask live, in the posting or in the application form?
In the form, as one visible field, with the question also stated in the posting so nobody starts an application they did not want. One field, one question, a stated time budget. Do not stack it with an essay box, a second cover letter and a set of screening questions the ATS added by default: the point is a single deliberate cost, not a longer path. Read every answer, or remove the field.
What if the ATS will not host a free-text question?
Put the question in the last line of the posting and ask for the answer in the first line of the cover letter field, which every system has. A dedicated email alias works too, though it splits your record of who applied. Either way, decide in advance who reads the answers and when, because an ask nobody reads is worse than no ask: it costs the candidate the fifteen minutes and returns nothing.
How long before a posting change shows up in the numbers?
One full cycle, so two to three weeks for most roles, and longer if the posting is syndicated to boards that cache it. Compare the same posting with itself, before and after, and count reads rather than applications. If the only thing that moved is the raw count, the change was cosmetic. If reads per hundred applications moved, the targeting worked.
References
- 1. GPT detectors are biased against non-native English writers pmc.ncbi.nlm.nih.gov Supports the claim that applications cannot be sorted by whether a model wrote them: seven detectors averaged a 61.3% false-positive rate on 91 human-written TOEFL essays.
- 2. California Labor Code Section 432.3 leginfo.legislature.ca.gov Supports the claim that pay-scale disclosure in a posting is already law in California for employers with 15 or more employees, including postings placed by third parties.
- 3. Revisiting Meta-Analytic Estimates of Validity in Personnel Selection: Addressing Systematic Overcorrection for Restriction of Range gwern.net Supports the work-sample and structured-interview estimates used to justify a short role-shaped ask: work samples at .33, revised down from .54, and structured interviews at .42.
3 sources, numbered by first appearance. 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.