Why Did Applications Per Opening Triple, and How Many Are Real?

Why did applications per opening triple this year?

Because the cost of producing one collapsed while the number of openings didn't move. A tailored resume and cover letter used to cost a candidate an hour; a model now does it in twenty seconds, and one-click apply removes the upload. LinkedIn's own competition measure has job seekers sending roughly twice as many applications as before the pandemic, against a job-to-seeker ratio back near late-2019 levels ¹.

That is a per-seeker figure, and your inbox counts something else. Three multipliers sit between the two, and they compound:

• Volume per seeker. Roughly double, on LinkedIn's measure ¹. This is the only one of the three with a public number attached.
• Breadth per seeker. The same candidate now applies across adjacent titles, adjacent seniority and adjacent cities, because the marginal cost of one more application is a click. A posting that used to see only its own obvious candidates now sees the near-misses from four other searches.
• Distribution. Aggregators scrape and re-list, job-alert emails fire on title keywords, and a remote-eligible posting enters a national pool rather than a metro one. Nothing about the role changed; its audience grew by an order of magnitude.

Multiply a doubling by a widened net and a national audience and a tripling at the posting level is ordinary arithmetic, not an anomaly. It also means the national average is useless to you: it averages a remote analytics req that went up fivefold with a licensed on-site role that didn't move.

The usual first response is to tighten the automated filter, which fails for a separate reason: every application now contains every keyword in the posting, because the model that wrote it read the posting. If your ATS keyword screen stopped separating anyone, that is why, and a stricter threshold makes it worse rather than better.

How many of those applications are real?

Nobody has published a credible number, and a vendor quoting one is selling something. What the evidence supports is narrower: the bulk of the increase is real people applying less selectively, not synthetic applicants. AI-written material is not a proxy for fake either. In a field experiment across nearly half a million job seekers, algorithmic writing assistance raised hire rates by about 8%, with no drop in employer satisfaction afterwards ².

"Real" is doing too much work as a word. Split it into three questions you can actually answer, in ascending order of cost:

1. Is there a person behind it? Almost always yes. Identity fraud in applications is a genuine problem in remote-only, contract and offshore-eligible roles, and it is handled with identity verification at offer, not with resume reading. It is the smallest of the three categories and the only one that belongs to security rather than to recruiting. 2. Do they meet the hard requirements? Checkable in seconds if you ask for the checkable things: work authorization, licence number, the named system, the location they can actually work from. This is where most of the new volume fails, and it fails quietly because nothing in the application says so. 3. Would they take the job? The expensive one, and the one that determines whether your funnel math holds. A candidate who applied to four hundred roles has no particular relationship to yours, and will behave accordingly at the scheduling stage.

The practical consequence is that "how many are fake" is the wrong question and "how many are addressable" is the right one. A serious candidate who used a model to write a clear resume is not a degraded applicant. An AI-written resume is not a reason to reject anyone, and treating it as one removes exactly the people who use their tools well.

How do you measure intent density instead of guessing?

Pick one number and compute it weekly. Intent density is the share of applications that clear your hard requirements and answer two short intent questions consistently. Put the questions on the application itself, keep them to one screen, and tag every application with the source it arrived from. Two weeks of that tells you more than any estimate of what fraction of the market is spray.

Four fields do the work, and none of them are essays:

• Location and work authorization, as a constrained choice rather than free text. Not a filter, just a fact you now hold.
• Compensation expectation, as a number in a band. Where a pay range is posted, a figure well outside it is a candidate who did not read the posting.
• One role-specific question, answerable in two sentences by anyone who has done the work and unanswerable from the posting alone. "Which system did you run the month-end close in, and what broke most often?" A model will produce something; a practitioner produces something specific.
• A single yes/no on the actual constraint: the on-site days, the shift, the travel, the clearance.

The formula is one line: applications that clear all four, divided by applications received, cut by source and by week. The absolute level means nothing and there is no benchmark worth quoting. The spread is the signal. When one source runs at four times another's density, you have found where to spend and where to stop, and you found it with your own data rather than an industry average.

Run the same cut on time-to-first-response, because density and responsiveness usually move together. Then stop reporting raw application counts to hiring managers entirely; a number that tripled for reasons unrelated to your role's attractiveness is a number that misleads everyone who sees it.

Which postings absorbed most of the increase?

Remote-eligible roles with ordinary titles, in functions the market reads as generalist: analytics, marketing, product, program management, customer success. A location-bound role with an unusual title and a licence requirement barely moved. Public breakdowns by function are thin, so treat this as a shape to check against your own postings rather than a statistic. The check takes ten minutes and settles it.

Three diagnostics separate a posting that tripled from one that didn't:

• Applications in the first 72 hours as a share of the total. A front-loaded curve means alert-driven, title-matched traffic. A flat curve means people who went looking for your role specifically.
• Share arriving outside the posted geography or pay band. This rises with remote eligibility and with title generality, and it is the cleanest proxy for how much of the increase is breadth rather than interest.
• Share from aggregators versus your own careers page. Direct applicants read the posting. Aggregator applicants often saw a title and a company name.

Title wording moves this more than most recruiters expect. A common title collects everyone whose alert matches it; a specific one collects fewer people who are closer. The same is true of requirement lines. An AI requirement written vaguely invites the whole market, and writing the AI requirement precisely is one of the few levers that reduces volume and raises quality at the same time.

If your role is in the absorbing group, plan for the volume rather than fighting it. The filtering approach that worked at eighty applications does not scale to four hundred, and the honest options are a shorter checkable filter plus an earlier work sample, or assessing everyone who clears the hard requirements and skipping the reading entirely.

Should you filter harder or assess earlier?

Assess earlier. A harder filter on a pool this size mostly removes people you cannot read, and any rule you use to decide who advances is a selection procedure: if it screens out a protected group at a higher rate, you have to show it is job-related and consistent with business necessity, and you carry that obligation even when a vendor supplied the tool ³. A keyword threshold or a judgment about tone cannot carry it.

Work samples carry the weight instead. Tasks that resemble the job's tasks closely enough that performance on one tracks performance on the other hold up as evidence, and they generally show little or no performance difference by sex or race; candidates also tend to accept them as fair, because the connection to the job is visible ⁴.

The cost objection is real and the arithmetic usually survives it. Four hundred applications at ninety seconds of reading each is ten hours. A checkable filter on work authorization, location and one licence or system requirement cuts that pool hard on facts rather than on prose, and the hours you save fund a real task for the people who clear it. Where the volume still overwhelms that, dropping the resume screen and going straight to a work sample is cheaper than it sounds.

One caution about the task itself. A take-home a model finishes in thirty seconds measures nothing, so the exercise has to allow AI openly and be built so that how the candidate uses it is the thing being read. That is a design constraint, not a detection problem, and it is the difference between an assessment that survives this volume and one that quietly stops discriminating between anybody.

See a sample report >