Indeed Matches Every 2.2 Seconds. Robert Half Says Hiring Slowed 2 Weeks.

Two numbers from 2026 look like they belong to different planets. Indeed says its AI produces a job match every 2.2 seconds. Robert Half says 20% of hiring managers are now two weeks slower because of the resumes those matches produce. They are the same number, measured at opposite ends of the same broken pipe.

If you run hiring at a startup or a mid-market company right now, you already feel this. The inbox is fuller. The signal is worse. The best candidate you have talked to this quarter came from a warm intro, not a job board. This piece is about why that is structural, not seasonal, and what actually works in the AI-resume flood.

The 2.2 seconds is the two weeks

Recruit Holdings, Indeed's parent, has been repeating the "a job connection every 2.2 seconds" line since a July 2025 internal memo. In May 2026 they updated the framing to 31 hires per minute across 665 million Job Seeker Profiles and 3.5 million employers. It is a genuinely impressive throughput number if you believe throughput is the thing you want.

Robert Half surveyed more than 2,000 U.S. hiring managers in late 2025 and published the results on March 10, 2026. Two thirds of HR leaders said AI-generated applications were slowing their hiring. One in five said the delay was more than two weeks. Eighty-four percent said their workloads got heavier. Sixty-five percent said AI-enhanced resumes made skills harder to verify. Dawn Fay, Robert Half's operational president, described "a surge in unverified applications extending timelines and delaying critical work."

Both things are true simultaneously because they describe the same event from different sides. Every automated match on the platform side is an unqualified application on the employer side. The platform metric and the employer metric are the same wire, measured at each end.

11,000

LinkedIn confirms the scale. The platform is processing roughly 11,000 applications per minute, up more than 45% year over year. HR consultant Katie Tanner pulled a single remote posting after it drew more than 1,200 applications and spent three months triaging what was left. She is not an edge case. She is the median employer with a public URL.

Indeed is quietly conceding the point

Look at what Indeed shipped in June 2026. A "Sourcing Assistant." Marketed with survey data from its own Harris Poll: 71% of hiring managers say higher application volume has made it harder to find qualified people, 72% fear missing top talent in the pile, and 93% have already lost top talent because hiring took too long.

Read that again. The world's largest job board is now selling an outbound sourcing product against its own inbound funnel, using its own survey to argue that the inbound funnel is drowning employers. Its own technical documentation says the new matching "goes beyond matching keywords on resumes or job descriptions." That is a public admission that keyword scoring, the foundation of ATS filtering and job-board relevance ranking for the last twenty years, no longer works.

The character detail worth holding onto: Recruit Holdings cut about 1,300 people, roughly 6% of its HR-tech segment, in July 2025 while doubling down on AI matching. CEO Hisayuki "Deko" Idekoba has talked publicly about hiring as a two-sided "yes" problem. He is running that thesis by removing the humans on his side of it. The platform selling matches is shrinking the humans behind those matches, which is a coherent bet if you believe the machine can do it and an incoherent one if you believe Robert Half's number.

Why the AI vs AI arms race is a dead end

The obvious response to AI resumes is more AI screening. Score harder. Verify with a model. Ask the LLM to catch the LLM. Every ATS vendor is shipping some version of this pitch in 2026.

It does not work, and the reason is structural. Keyword-based filtering optimizes for exactly the signal that AI-tailored resumes are engineered to produce. When a candidate pastes your job description into ChatGPT and asks for a resume, the output is a document that maximizes the score your ATS will give it. You are grading a test the other side wrote the answer key to. Raising the passing grade does not help. You get more high-scoring documents and the same underlying skill distribution.

You are grading a test the other side wrote the answer key to.

The July 2026 "Who is hiring?" thread on Hacker News is where you can see the smarter half of the market reacting. Real listings this month include playit.gg, which asks candidates to email careers@playit.gg directly with no ATS in the loop. AveryIQ writes, in the actual post, "under no circumstances send any AI-generated content." Surge AI, Foxglove, Mentimeter, and Pango (YC S26) are running similar direct-contact plays. This is not a Luddite reaction. It is a sourcing strategy. These founders are opting out of the 11,000-per-minute funnel entirely and reverting to small-batch, high-intent, human-authored channels.

The only asymmetry left is intent

If keyword scoring is dead and AI-vs-AI is a treadmill, what remains? The one thing an LLM-tailored resume cannot fake in bulk: what someone has actually built, shipped, said, and been credited for in public.

That means sourcing starts from the question, not the corpus. Instead of "whose resume contains Rust and Kubernetes," you ask "who has actually shipped a production Rust service on Kubernetes in the last eighteen months, ideally at a company smaller than 200 people." Instead of "who has 'staff engineer' in their headline," you ask "who has landed a merged commit in a top-100 Python package this year." The first query returns 40,000 resumes. The second returns 60 humans, and you can talk to all of them this month.

This is the thing job boards structurally cannot do. Their asset is the resume database. Their revenue depends on more resumes, not fewer, and the AI apply tools are giving them exponentially more. The 2.2 seconds number gets faster next year. So does the two-week delay.

This is the exact friction we built Refolk for. You describe the person in plain English, the way a hiring manager would describe them to a colleague, and you get a ranked shortlist pulled from GitHub, LinkedIn, and the open web. No boolean strings. No keyword bingo against a resume the candidate had a model write. The signal comes from what the person actually did in public, not what a document claims.

What this changes about the funnel

If you accept that inbound is broken for skilled roles, three things change about how you run hiring.

First, the top of your funnel gets smaller and better on purpose. You stop measuring applications per week. You start measuring "how many of the 20 best possible people for this role did we actually reach out to." That is a knowable number and it is almost always under 20% for the roles you care about most.

Second, your recruiter's job changes shape. Less time in the ATS queue triaging AI slop. More time writing outbound that a specific person will actually read because it references a specific thing they built. Volume metrics like "31 hires per minute" and "665M profiles" become anti-signals. They tell you how much noise the platform has, not how much signal you can get from it.

Third, the tooling stack inverts. The ATS becomes a system of record, not a system of discovery. Discovery moves to intent-based sourcing tools that read the open web the way a good sourcer would if they had a thousand hours per role. Refolk sits in that discovery layer, which is why the natural workflow is "ask Refolk in plain English, then push the shortlist into Ashby or Greenhouse," not the other way around.

The founders on HN already figured this out

Read the July 2026 "Who is hiring?" thread top to bottom. The good posts have three things in common. They name the specific product surface the hire will own. They give a direct human email. They ask for human-written signal (a paragraph, a project link, a story) rather than a resume.

The pattern is not nostalgia. It is a rational response to the fact that a resume in 2026 is a low-information artifact. The candidates who care are giving founders a way to prove they care back. The founders who care are giving candidates a way to prove they are not one of 1,200 auto-applications. Everyone in this exchange is opting out of Indeed's 2.2 seconds because that number is the problem, not the product.

The rest of the market will follow, because the math forces it. If your competitor is hiring by naming a specific person who shipped a specific thing and emailing them, and you are hiring by ranking 4,000 tailored resumes, you lose the offer, then you lose the ship date, then you lose the quarter. Robert Half's two-week number is what that loss looks like in aggregate. It gets worse from here unless you change what you are measuring.

What to do this quarter

Three concrete moves, in order.

Pick one role and run it entirely outbound for 30 days. No job board post. Define the person in one sentence a hiring manager would say out loud. Source 40 to 80 named humans from that sentence. Reach out to all of them personally. Track reply rate and first-conversation quality, not applications. This is the baseline you need to know whether inbound is actually giving you anything for that role.

Audit your last 10 hires. For each, write down the true first-touch channel: referral, outbound, inbound. For the inbound hires, ask honestly whether they would have shown up in an outbound search. Most of them would have. That is your permission slip to shift budget.

Stop treating the ATS pile as the source of truth. It is a queue of documents optimized against your own filters. The people you actually want to hire are visible in GitHub commits, conference talks, package maintainer lists, WARN notices, and last week's HN thread. Sourcing tools like Refolk exist to read those surfaces the way a senior sourcer would, at speed, from a plain-English brief. That is the asymmetry that survives the AI-resume flood, because it does not depend on the resume at all.

Indeed will keep announcing a faster number. Robert Half will keep publishing a longer delay. Both numbers are true. Neither one is going to help you close your next staff engineer. What closes them is a specific message, sent to a specific person, based on something they actually built. Everything upstream of that is now noise.

FAQ

Is Indeed's "31 hires per minute" number wrong?

No, it is almost certainly accurate for what it measures: a hire event attributed to the platform anywhere in the funnel. The issue is that throughput at that scale is precisely what generates the volume problem employers now report. Indeed's own June 2026 Harris Poll found 93% of hiring managers have lost top talent because hiring took too long. Both statements come from the same company in the same quarter. The throughput is real and it is also the constraint.

Are AI-generated resumes actually that easy to spot?

Not reliably, and that is the point. Robert Half found 65% of HR leaders say AI-enhanced resumes make skills harder to verify. Detection tools have poor precision and false-positive rates that will get you sued. The durable answer is not detection. It is sourcing from public evidence of work (commits, papers, launches, patents) so that the resume becomes a formality after the conversation, not the filter before it.

Does this mean we should stop posting jobs entirely?

For commodity roles with abundant qualified supply, job posts still work. For any role where the top 50 candidates matter more than the next 5,000, treat the job post as a passive backup and run the role outbound. The HN founders naming direct emails, like playit.gg and AveryIQ, are doing exactly this. Post the role, but hire from the shortlist you built by asking who has actually done the work.

How is intent-based sourcing different from a better boolean?

Boolean queries search text on resumes and profiles. Intent-based sourcing starts from a description of the person and their work, then pulls evidence from wherever that evidence lives: GitHub, LinkedIn, conference sites, package registries, patent databases, the open web. A boolean returns everyone whose profile contains "LLM inference." An intent query returns the 60 people who have actually shipped LLM inference infrastructure at a company under 300 employees in the last year. The second list is the one you can actually work through.