Evidence audit Β· Hiring funnel Β· August 2026
What the large-sample data actually says about getting hired β which tactics have real effect sizes behind them, which are marketing copy with no methodology, and where the leverage genuinely is.
Almost everything written about job-search tactics is uncited. The "75% of resumes are rejected by an ATS" figure traces to a 2012 sales pitch by a company selling resume optimization. The "90% of interviews go to people who apply within 24 hours" figure has no published methodology anywhere. Both are load-bearing beliefs for millions of people and neither has a study behind it.
What does hold up, across hundreds of millions of applications: the channel you arrive through swamps every other variable. A referred candidate converts at roughly 11Γ an inbound applicant; an internal transfer at 32Γ. Nothing you do to a resume, a cover letter, or your submission timestamp comes within an order of magnitude of that.
So the data-driven answer to "how do I get a high-paying job" is not a trick. It is: stop optimizing the 0.5% channel and move to the 5% channel. Everything below is the case for that, plus the honest error bars on each piece.
Every claim in this document carries one
A number is only as good as the design that produced it. Grades here describe evidence quality, not how much I like the conclusion.
| Grade | Means | Example |
|---|---|---|
| A | Randomized experiment, or platform census with n in the tens of millions | Resume-writing RCT, n β 480,000 |
| B | Large observational dataset, credible source, but confounded β no randomization | Referral conversion rates from ATS vendors |
| C | Small sample, self-report, or a vendor measuring something it sells | Salary-negotiation surveys; the 7.4-second study |
| D | No methodology published, or traceable to marketing material | The "75% ATS rejection" figure |
You asked for 100% certainty. That does not exist in this domain and I would be lying to offer it. What exists is a set of effect sizes with wildly different confidence attached, and the useful skill is telling them apart.
Gem Β· 165M applications Β· 1.2M hires Β· Ashby Β· 54M applications
Start here, because every tactic below is a multiplier on this number and a big multiplier on a tiny number is still a tiny number.
And the denominator is getting worse fast, for a specific and recent reason:
Greenhouse's CEO calls this the doom loop: candidates use AI to apply at scale, recruiters respond with AI screening, rejected candidates escalate to more automated applications. In a 2026 ZipRecruiter survey of 1,000+ talent-acquisition professionals, 48% said AI has increased application volume per role and 92% reported AI somewhere in their hiring process. B
The strategic consequence matters more than the statistic: volume tactics have been commoditized to zero. Whatever edge existed in applying to many jobs quickly has been competed away by $20/month tools. Any strategy that a bot can execute is now worthless by construction.
The honest answer is smaller than the internet claims
I went looking for the studies behind the timing claims. They do not exist. Here is what I found instead.
"90% of interviews go to people who apply within 24 hours." Appears across dozens of career sites with no primary source, no sample size, no methodology. It is also arithmetically strange: postings collect only about half their applications in the first week, so this would require a near-total collapse in conversion afterward.
"Applying in the first 24 hours makes you 8Γ more likely to get an interview." Same problem. No traceable study.
"Day 1 applications get 12% response, Day 4 gets 61%." A self-tracked set of 347 applications posted to social media. Uncontrolled, unverifiable, and a 61% response rate sits ~15Γ above the population interview rate. Discard.
LinkedIn's "first 25 applicants are 3Γ more likely to get hired." This one is real in the sense that LinkedIn published it β but it is a correlation from a company selling job-alert engagement, and it is heavily confounded. People who apply in the first 25 are disproportionately those with alerts configured, who already follow the company, who are actively and seriously searching. They were more likely to be hired for reasons that have nothing to do with the clock.
Now the part that is measured, and it does support acting fast β just less dramatically:
For candidates who never reach an interview, the median time to archive is 6 days, and 58% are archived within 9 days. The decision window on a given posting is roughly one business week. Miss it and your application is being read, if at all, against a pool that has already produced finalists.
But Ashby's own product documentation notes that the review queue is configurable in both directions β recruiters can work newest-first or oldest-first. In a newest-first queue, applying early actively buries you. There is no universal first-mover advantage because there is no universal review order.
Apply inside 72 hours because it is free, not because it is transformative. The plausible effect is a modest relative lift on a sub-1% base rate β moving you from perhaps 0.4% to 0.6%. That is a real improvement and it costs you nothing, so take it. It is not a strategy. If your plan is "apply faster," you have chosen the cheapest lever in the entire system and it will not get you into OpenAI.
On the literal question β will the employer check my email? β if you applied through a portal there is no email. Your submission joins ~291 others in a queue and receives somewhere between 10 seconds and 2 minutes of human attention, or none, within about six days. If you send an actual email to an actual person, that is a completely different question with a completely different answer, covered further down.
Where the 75% figure came from, and what actually rejects you
The claim that applicant tracking systems auto-reject three-quarters of resumes traces to a 2012 sales pitch by a company called Preptel β which sold resume-optimization services and has since gone out of business. No methodology was ever published. It has been repeated for fourteen years because it is useful to everyone selling a fix for it. D
Investigations of the major systems find that none of them automatically discard or hide resumes by default. A 2025 survey found ~92% of recruiters do not configure content-based auto-rejection β though note that survey had n = 25, which is far too small to lean on. C The stronger evidence is simply the vendor documentation: the feature people fear is not the default behavior in these products.
Two things do genuinely filter you out automatically, and they are worth separating from the myth:
So resume optimization has a real but bounded effect, and we can put a number on the ceiling, because someone ran the experiment.
Roughly 480,000 jobseekers on a large online labor market were randomized to receive algorithmic writing assistance on their resumes or not. Treated jobseekers saw an 8% increase in the probability of being hired, with no evidence that employers were less satisfied with them afterward. The proposed mechanism is that better writing helps employers assess ability rather than signaling ability itself. A
Read that carefully, because it is the single most useful calibration in this document. A genuinely good rewrite of your resume β the thing an entire industry exists to sell β buys about 8% relative. Applied to a 0.5% offer rate, that is 0.54%. It is worth doing exactly once. It is not worth doing eleven times, and it is certainly not worth the weeks people spend on it.
One adjacent finding on what recruiters actually look at: the widely cited "recruiters spend 7.4 seconds on a resume" figure comes from a 2018 eye-tracking study of 30 recruiters, run by a job board. C Small and commercially motivated β but the qualitative point survives: that 7.4 seconds was never reading time, it was the decision about whether to keep reading.
This is the finding. Everything else is a rounding error against it.
The same pattern shows up everywhere it has been checked, which is why I put weight on it despite the confounding:
Ashby splits referrals by who made them, and the gap is large:
| Referral source | Reach interview | Get hired |
|---|---|---|
| Referred by someone in the same function | 37% | 5.2% |
| Referred by someone in a different function | 26% | 3.1% |
| Inbound application (no referral) | ~4% | ~0.5% |
A referral from an engineer on the team you want is worth about 1.7Γ a referral from a random employee β and roughly 10Γ an inbound application. Business roles shown; technical roles convert lower throughout the funnel, so treat 5.2% as an optimistic bound rather than your expected value.
This changes the target. Most people ask recruiters or whoever they can find at the company. The data says: find the people who do the specific job you want, on the specific team you want. A referral from a designer will not help your infrastructure application nearly as much.
The causal evidence is thinner but points the same way. Burks, Cowgill, Hoffman & Housman (Quarterly Journal of Economics, 2015) studied personnel data from nine large firms across three industries and found referred applicants are more likely to be hired and to accept offers despite similar measured skills β and that referred workers are 10β30% less likely to quit and outperform on rare high-impact outcomes like patents. B Firms prefer referrals because referrals actually work for them, which is why the channel is durable and not an exploitable loophole that will close.
Note what this does and does not say. It does not say referrals are easy β 43 warm referrals is an enormous amount of relationship work. It says that if you are going to spend six months on this, spending it on 43 relationships strictly dominates spending it on 459 applications, and the 459-application path is the one most people take.
Lower than the guides promise, high enough to be the best available substitute for a network
If you have no network, cold outreach is how you manufacture one. Here are the benchmarks, and one important caveat about direction.
| Metric | Rate | Source & grade |
|---|---|---|
| Cold email reply, all sectors | 3.4% | Instantly.ai, billions of emails B |
| Top quartile senders | 5.5% | Instantly.ai B |
| Top decile senders | 10.7% | Instantly.ai B |
| Recruiting outreach, 4+ touch sequence | 6β8% | Aggregated vendor data C |
| Generic, untargeted sends | 1β2% | Aggregated vendor data C |
| "Cold emails to hiring managers" | 15β25% | Content-marketing sites, no methodology D |
The direction caveat matters. Nearly all published outreach data measures recruiters emailing candidates β a high-status party offering something to a lower-status one. You are running that arrow in reverse, asking rather than offering. Your realistic reply rate is at or below the general benchmark, not above it. Plan on 3β8%. Fifty well-targeted emails buys you two to four conversations.
Two findings from Gem's analysis of 4 million recruiting emails are worth copying regardless of direction: B
No RCT exists. Here is the honest reasoning anyway.
You asked specifically about writing a custom email with ideas about a product a company is building. I could not find a study that measures this, and I am not going to invent an effect size for it. D on direct evidence. But the surrounding evidence is informative in two directions, and one of them may surprise you.
Sackett and colleagues' 2022 reassessment in the Journal of Applied Psychology re-ran the foundational meta-analyses of personnel selection after correcting for systematic overcorrection for range restriction. A Structured interviews came out strongest (r = .42), ahead of cognitive ability (r = .31) β and work samples were revised downward. So an unsolicited project is not some superior predictor that employers are irrationally ignoring. Companies are not underrating your side project; they have reasonable evidence that structured interviews tell them more.
There is also a specific failure mode. An outsider's "here's how you should build your product" email is, from the receiving side, usually a confident proposal from someone who lacks the context that makes the actual problem hard. That reads as a negative signal about judgment, not a positive one about initiative. Cold outreach to a large AI lab already skews toward this genre, which is part of why it gets ignored.
The mechanism a project actually exploits is not predictive validity β it is search cost. In a 291-applicant pool getting 10 seconds to 2 minutes each, the binding constraint is attention, not assessment. An artifact that can be evaluated in under a minute and is unmistakably about this team's problem changes which pile you are in. That is a real and defensible reason to build something, and it is a different reason than "it proves I'm good."
Ashby's take-home data is the closest empirical analogue: candidates who complete take-home assignments convert to hire at 13% overall, and up to 22.7% in high-volume business roles (7.6β14.3% technical). B That is heavily confounded β you only reach a take-home after passing screens β but it establishes that once evaluative work enters the picture, conversion rates look nothing like 0.5%.
So: build something, but constrain it hard.
Process, compensation, and the one documented channel
Their published hiring process runs: recruiter screen β role-specific technical screen β a paid work trial or take-home β a 4β6 interview loop covering technical depth, system design, and mission alignment β offer. Four to eight weeks end to end. C β this is assembled from interview-prep sites, which are commercially motivated and not primary sources; treat the shape as reliable and the details as approximate.
Compensation, from Levels.fyi as of 26 August 2026 B β self-reported but verified against offer letters and W-2s, with a known upward selection bias since people with strong offers are likelier to submit:
| Level | Total compensation | Note |
|---|---|---|
| L2 | $253K | Entry |
| L3 | $333K | $213K base + $118K stock |
| L4 | $644K | $269K base + $375K stock |
| L5 | $916K | $326K base + $589K stock |
| L6 | up to $1.52M | Thin sample at this level |
| Median, all SWE | $875K | Anthropic comparable: ~$367Kβ$1.25M |
Two things worth knowing about the front door.
The residency is a genuinely different channel. OpenAI runs a six-month paid Residency at roughly $220K annualized, and its stated target is explicitly non-traditional: self-taught engineers, independent builders, and researchers from mathematics, physics, and neuroscience without ML backgrounds. A on the terms β this is from OpenAI's own site. That is a company publicly announcing that it will evaluate people who cannot compete on a conventional resume. If your background is unconventional, this is a materially different funnel from the standard SWE posting, and most people do not consider it.
On cold applications there. The instructive documented case: an engineer with director-level experience at Microsoft and Meta cold-applied to OpenAI in 2022 and received nothing β not even a screening call. A year later, a former Meta colleague who had joined referred him, and he got in. There are also documented cases of people landing OpenAI roles through cold LinkedIn DMs. D β these are two anecdotes and prove nothing on their own. I include them only because they are directionally consistent with the 11Γ referral multiplier that is well-measured, and because the first one describes a candidate far stronger than most who still got nothing from the front door.
Effect size Γ feasibility Γ· time cost
| # | Action | Best estimate | Evidence | Time cost |
|---|---|---|---|---|
| 1 | Get a referral from someone in your target function on your target team | ~10Γ vs inbound | B 165M apps | Daysβweeks each |
| 2 | Negotiate every offer you receive | +12β19% | C surveys | ~2 hours |
| 3 | Cold-email individual engineers, 4-touch sequence | 3β8% reply | B 4M emails | ~20 min each |
| 4 | Clear the knockout questions β work auth, location, YoE | Binary gate | B vendor docs | 5 min |
| 5 | Write one genuinely good resume β then stop | +8% relative | A RCT nβ480K | One-time |
| 6 | Build a narrow artifact for your top 5 targets | Unmeasured | D mechanism only | 10β40 hrs each |
| 7 | Apply within 72 hours of posting | Modest, confounded | C | Free |
| 8 | Keyword-align each resume to the posting | Unmeasured | C mechanism real | 10 min each |
| 9 | Mass-apply using AI tools | Net negative | B | β |
On #2, which people skip. Roughly 66% of workers who negotiate get what they asked for, and about 55% never try; reported average increases run 12β19%. C β survey data with obvious self-report bias, so discount it. But do the arithmetic anyway: at the compensation bands above, a 12% improvement is well over $100,000 per year for two hours of preparation. Nothing else in this document has a comparable return per hour, and it is the one item that is fully within your control.
On #9. Automated mass-application is what produced the 291-applications-per-hire environment. It is now actively harmful β Greenhouse specifically flags $20 mass-apply tools as the source of the spam recruiters are filtering against, and getting caught in that bucket is worse than not applying.
The most important methodological point here
You want to be driven by data, so this needs saying plainly. Suppose you want to know whether your new resume beats your old one, and suppose it is genuinely, substantially better β a 50% relative improvement, lifting your interview rate from 4% to 6%. To detect that at conventional statistical power:
You will never send 3,720 applications. Therefore you can never personally verify which of your tactics work. Any conclusion you draw from your own 40 or 80 applications is noise β the variance at these base rates completely swamps any real effect you could produce. This is precisely why the job-search advice industry can sustain thousands of contradictory confident claims: nobody, including the people making them, can falsify any of it from personal experience.
This has one clean implication. Since your own funnel cannot generate reliable evidence, you must borrow population-level evidence and follow it even when it feels wrong β and that evidence says the channel dominates everything else. Track your own numbers for process discipline and morale, not for inference. Log applications, channel, date, and outcome; use it to notice that you've sent nine cold emails this month instead of the forty you planned. Do not use it to conclude that cover letters work.
And one honest closing note on the framing you opened with. There is no set of actions here that produces a job with 100% certainty, and the strength of this evidence is not remotely uniform β one randomized trial, several very large observational datasets, and a great deal of confident nonsense that I have tried to label as such. The realistic goal is not certainty. It is to stop spending effort in a channel with a 0.5% conversion rate when a 5% channel is available to you, and to be honest with yourself about which one you are actually working in this week.