Gem's 2026 Report Says Mine the ATS. The Math Says Don't.
Gem's 2026 benchmarks show 46% of sourced hires now come from ATS rediscovery. That's a top-of-funnel confession, not a sourcing strategy.
Gem's 2026 Recruiting Benchmarks Report dropped with one headline number every TA leader is now quoting: 46% of sourced hires come from rediscovered ATS candidates, up from 26% in 2021. Inside the report, that's framed as a productivity win. Read with the rest of the numbers, it's an admission that the top of the funnel collapsed and recruiters are eating their own archive to make quota.
The same week, Elly shipped its AI Sourcer (June 2, 2026), pitching conversational search across "proprietary data sources." Every AI sourcing vendor is converging on the same idea: mine the closed pools you already own. That's a convenient story for software companies that sell ATS layers. It's a worse story for anyone actually trying to hire senior engineers in 2026.
What the 2026 numbers actually say
Gem analyzed 165M+ applications, 15M candidates, and 1.2M hires for this report. The top-line stats are brutal, and they all point in the same direction.
Recruiters are handling 93% more applications than in 2021. They're managing 40% more open roles (13.4 reqs on average). TA teams are 14% smaller. Hires per recruiter dropped 43%. Interviews per hire are up 33% since 2021, on top of the 2025 baseline of 20 interviews per hire and a 41-day time-to-hire.
You can read those numbers two ways. Gem's reading: AI rediscovery is unlocking productivity, and the proof is that 46% of sourced hires now come from your existing database. The other reading: when you take a 14% smaller team and double its inbound load, the only sourcing that still fits in the day is the kind that doesn't require leaving the ATS.
That's not a strategy. That's a triage pattern.
The "silver medalist" myth needs to die
The pitch for ATS rediscovery has always leaned on the silver medalist: the candidate who almost got the offer last cycle, ready to convert this cycle. It's a clean story. It's also mostly fiction at the volume Gem is describing.
A candidate who lost out 18 months ago has a stale skill profile, a known comp anchor, and a recorded "no" in their history (yours or theirs). At 46% of sourced hires, you are no longer skimming the genuine near-misses. You are scraping people who applied to a different role, at a different stage, against a different bar, and calling that sourcing.
Meanwhile, Gem's own data says sourced candidates convert 8x better than inbound. Referrals convert at 11x. Internal mobility at 32x. Job boards generate 90% of applications and roughly half of hires. The report tells you exactly where the leverage is, then sells you a feature that operates on the lowest-leverage pool in the funnel.
Rediscovery is what you do when you no longer have time to source. It is not what sourcing is.
The actual silver medalists in 2026 are not in your Greenhouse archive. They're the staff engineer who shipped a high-velocity PR to tokio last week, the maintainer who just gave a talk at a ClickHouse meetup, the infra lead whose name appears three times in the CockroachDB contributor graph this quarter. They never applied to your company. They never will, unless someone reaches out with a reason.
The 30% of engineers your ATS will never see
Here is the structural problem with ATS rediscovery and LinkedIn-shaped AI sourcers (Gem, Elly, SeekOut, hireEZ, and most of the rest): roughly 30% of software engineers are not on LinkedIn at all. The skew is worst exactly where senior teams hurt most: infra, systems, distributed databases, compilers, OSS-heavy stacks.
If a third of the market never built a LinkedIn profile, and the people who applied to your company are a self-selected slice of the two-thirds who did, your ATS is a sample of a sample. No amount of conversational search on top of that pool widens it. It just makes the same closed graph easier to query.
GitHub had 100M+ developers and over 1B contributions as of 2024. The GitHub Archive captures 30M+ public events per month: every commit, PR, review, and release, fully queryable. That is the open-web equivalent of an ATS, except it covers the entire industry and the data is verifiable. Most mainstream TA stacks do not touch it.
This is the gap Refolk was built around. You describe the person in plain English ("Rust async maintainers in the US who have shipped to tokio or hyper in the last 12 months and have spoken at a conference") and get a ranked shortlist drawn from GitHub, LinkedIn, and the open web together. The point is not to replace your ATS. The point is to stop pretending your ATS is the market.
Why VCs already moved here and recruiters haven't
SignalFire publishes an "Open Source Superstars" Top 100, ranking U.S. OSS contributors by ship velocity, maintainer impact, and ecosystem pull. The H1 2025 edition pulled from GitHub events between May and mid-August 2025. It is a public hiring map produced by a top-tier VC.
That should embarrass the recruiting industry. VCs, who do not hire engineers at scale, treat the open contribution graph as the canonical talent layer. The largest TA vendors, who do hire engineers at scale, treat it as out of scope and sell another pass over the ATS instead.
The emerging GitHub-signal layer (Fonzi, Kula, daily.dev Recruiter, Skillsync, TalentSeeker, Riem.ai) exists because the gap is obvious to anyone outside the ATS suite category. Vendor blogs in that space report personalized outreach referencing specific GitHub contributions at 25 to 40% response rates, vs. 10 to 15% for LinkedIn InMail. Treat those numbers as directional, but the direction is correct: specificity beats volume, and specificity requires signal the ATS does not store.
The interview load is a signal problem, not a process problem
The 33% jump in interviews per hire since 2021 gets framed as process bloat or candidate hesitation. It is mostly neither. It is what happens when the top of the funnel goes low-signal. If you can't tell from sourcing whether a candidate has actually built the thing you need built, you compensate downstream with more loops, more take-homes, more system design rounds, more references.
Better signal at the top collapses that load. A maintainer with three years of commits to a project in your stack does not need a coding screen to prove they can code. A speaker who explained your exact distributed-systems problem on a conference stage does not need a 90-minute system design round to prove they understand it. The cost of cheap top-of-funnel is expensive bottom-of-funnel, and Gem's own time-to-hire numbers are the receipt.
This is the second place Refolk earns its keep. When sourcing returns named people with verifiable artifacts (PRs, talks, blog posts, maintainer status), the hiring manager can pre-read before the first conversation, and the loop gets shorter because it was warranted in the first place.
What "rediscovery" should actually mean in 2026
There is a version of rediscovery that is genuinely useful, and it is not the version Gem's report sells. It is the version where you take the candidates already in your ATS, enrich them against current open-web signal, and surface the ones whose public work in the last six months matches what you are hiring for now.
That is a different operation. The ATS gives you the historical relationship. The open web tells you what they have actually been doing since. The intersection is small, high-quality, and not what conversational ATS search produces by default.
Practical shape of that for an engineering org hiring in 2026:
- Start every senior or staff search with an open-web pull first: contributors to the projects your stack depends on, speakers at the conferences your engineers attend, authors of the posts your team has shared internally in the last year.
- Cross-reference against your ATS. Anyone who shows up in both lists is a real silver medalist. Reach out with reference to their recent public work, not their old application.
- Treat the rest of the open-web pool as net-new sourcing. This is the segment ATS-only rediscovery is mathematically guaranteed to miss, and it is where the senior infra and systems engineers actually live.
- Measure time-to-hire benchmarks engineering teams care about (calendar days from first contact to signed offer for senior+), not application-to-hire ratios. The first number tells you whether your signal is good. The second just tells you whether your job posts get indexed.
For step 1 and step 3, the open web is the constraint. There is no internal database that solves it, which is the whole point. Tools like Refolk exist to make that pull cheap enough to run on every req, not just the ones you're desperate on.
The vendor convergence is the warning, not the answer
Gem's 2026 report and Elly's June 2 launch are not opposing positions. They are the same position. Both assume the right move in a constrained market is to extract more value from closed pools (your ATS, a LinkedIn-shaped graph) using AI. Both are selling to the same VP of Talent who is being asked to do more with 14% fewer people.
The contrarian move, and the one a small group of teams is already making, is to widen the input. Treat the GitHub contribution graph, OSS maintainer threads, and conference programs as a first-class sourcing layer. Use the ATS for what it is good at (warm relationships, prior context, internal coordination) and stop pretending it is the talent market.
The recruiting benchmarks 2026 cycle will be remembered as the year the industry doubled down on closed pools. The teams that quietly outperform the benchmarks will be the ones that didn't.
FAQ
Is ATS rediscovery actually bad, or just oversold?
It's a useful tactic that's being mis-sold as a strategy. Re-engaging a genuine recent near-miss is high-leverage. Mining 18-month-old applicants from unrelated roles to backfill a broken sourcing motion is not. The 26% to 46% jump in Gem's data is too large to be explained by better tooling alone; the simpler explanation is that recruiters who used to source outbound no longer have time, and rediscovery is what fills the gap.
What's the fastest way to start sourcing on open-web signal without rebuilding our stack?
Pick one search per week where you'd normally lean on LinkedIn Recruiter or your ATS, and run it against open-web signal instead. Specifically: identify the two or three open-source projects your stack actually depends on, pull their recent contributor lists, cross-reference against current employers, and reach out referencing the specific work. You don't need to rip out the ATS. You just need to stop letting it define the candidate universe.
How do we measure time to hire benchmarks engineering leaders actually trust?
Stop measuring application-to-hire, which is a function of your job-post distribution, not your hiring quality. Measure calendar days from first outbound contact to signed offer for senior and staff roles, and measure interviews per hire on that same cohort. If your signal is good at the top, both numbers drop. Gem's 33% jump in interviews per hire is the clearest evidence in the report that the industry's top-of-funnel signal is getting worse, not better.
Does any of this apply to non-engineering roles?
Some of it. The open-web argument is strongest for technical roles because the artifacts (code, talks, papers) are public and verifiable. For design, you get Dribbble, GitHub, and personal sites. For research, papers and citations. For sales and ops, the signal is sparser and LinkedIn is still the dominant graph, so ATS rediscovery weighs heavier in the mix. The core point holds everywhere: if your tooling can only see closed pools, you are sampling a sample.