72.4% of Recruiters Went Onsite. Your Python Pool Just Lost 68%.

Your final round is now a flight. That is the quiet consequence of the 2026 return to in-person technical interviews, and it is about to break Q4 pipelines that were built for a remote-national candidate pool. If your reqs were written assuming a candidate anywhere in the U.S., and your loop now assumes a candidate who can be in your lobby on 48 hours' notice, you have a sourcing problem that no ATS report will surface until it is too late.

Why 72.4% of recruiters went back to in-person in 2026

Overlay-based AI cheating tools got good enough, fast enough, that remote verification alone stopped being defensible. Gartner data (via Computerworld, Aug 26, 2025) puts 72.4% of recruiting leaders now conducting at least one in-person round specifically to combat candidate fraud. In-person interview requests jumped roughly 500%, from about 5% of roles in 2024 to about 30% in 2025. Google, Cisco, and McKinsey have all reinstated onsite rounds. Amazon now asks candidates to sign a no-unauthorized-AI pledge.

The driver is not hypothetical. Fabric analyzed 19,368 interviews between July 2025 and January 2026:

• 38.5% of candidates were flagged for AI-cheating behavior overall.
• 48% flagged in technical roles specifically.
• Cluely and Interview Coder together accounted for 45% of the cheating methods flagged.
• Flag rate rose from 15% in June 2025 to 35% by December 2025 across a 50,000-candidate Fabric dataset.

Cluely launched April 20, 2025 under the "Cheat on Everything" tagline, hit 70,000 signups in its first week, raised a $15M Series A, and by November had rebranded to a general "meeting assistant" with the same overlay under the hood. Interview Coder advertises 150,000+ users at $799 lifetime and claims a 65% success rate on coding rounds. Final Round AI claims 10M+ users. The DOJ has separately prosecuted schemes in which 300+ U.S. companies unknowingly hired North Korean IT operatives using stolen identities. This is now a security review, not just a quality review.

45%

The in-person snapback is really a return to three ZIP codes

Onsite rounds apply a silent geographic filter that invalidates any candidate who cannot reach a physical office on 48 to 72 hours' notice, and that filter lands unevenly on the map. In Refolk's index of professional profiles, there are roughly 100,069 U.S.-based software engineers (Software, Senior, Staff) who list Python as a skill, and their distribution is not close to uniform.

A Bay Area company that flips a Python req from remote-national to onsite has just made roughly two-thirds of its addressable candidates unreachable. Relocation friction and 48-hour scheduling windows filter out everyone not already local. NYC does slightly better on absolute count but proportionally worse: at ~8% share, going onsite in Manhattan drops the pool by more than 90%.

The mechanism matters. Recruiters are not losing candidates to cheating. They are losing candidates to geography, and blaming it on cheating.

The recruiter-to-engineer ratio quietly got worse

There is roughly one technical recruiter for every 4.7 Python engineers in the U.S., and the recruiters are clustered in the exact same three metros absorbing the onsite mandate. Refolk's index shows 21,424 U.S. technical recruiters and sourcers, concentrated in Austin, Seattle, and the SF Bay Area, which happen to be the largest onsite anchor cities (Databricks, Figma, Ramp, and Glean in SF; Snowflake and Blue Origin in Seattle).

That creates a specific bottleneck for Q4:

1. Secondary metros (Indianapolis, Portland ME, Boca Raton show up in the Refolk data) have real engineers but few local recruiters.
2. Hub-city sourcers are being asked to fill hub-city reqs from a shrunken hub-city pool.
3. The overflow work of surfacing candidates in secondary metros, who might relocate for the right role, falls on sourcers who do not know those markets.

Pipelines will not bottleneck on candidate supply. They will bottleneck on sourcer bandwidth to find the right thirty candidates for a Bay Area loop instead of the right three hundred nationally. If your workflow assumes an InMail blast to anyone with "Python" and "senior" in their headline, you are already behind. Describing the shape of the person you actually need, in plain English, and getting a ranked shortlist filtered by commute radius, is the exact gap Refolk closes.

Detection vendors flag behavior, not binaries

Even the best detection stack does not remove the need for the in-person round, because vendors flag interview behavior rather than fingerprinting cheat-tool binaries. Per Interview Coder's own writeup, no detection vendor fingerprints the Cluely binary. Fabric and Sherlock flag response-timing loops, gaze patterns, and multimodal adversarial signals. They are useful, but they are probabilistic.

That means "detection plus remote" is not a substitute for "geography-collapsed onsite." The in-person round is load-bearing, because:

• Fingerprint-level detection does not exist for the top overlays.
• Cluely's November rebrand to a generic "meeting assistant" removed the easiest heuristic (candidate mentions the tool by name on LinkedIn or Twitter).
• The 70,000-signup week-one population is now indistinguishable from legitimate note-taker users on the open web.
• Cluely's 2025 breach exposed 83,000+ user records, which is a hiring-signal problem, not a solution.

Recruiters are not losing candidates to cheating. They are losing candidates to geography, and blaming it on cheating.

Where the collapse hits hardest: engineering coding rounds

The 45% Cluely and Interview Coder share is specifically a coding-round problem, not a general interview problem, which means engineering reqs disproportionately lose sourcing radius versus sales, PM, or design. Coding rounds are the environment where overlay quality is highest and where the AI answer is closest to a passing answer. That is where in-person verification is doing the most work, and where the geographic collapse compounds hardest.

A rough tier of exposure:

• Highest exposure: IC engineering (SWE, ML, platform), where coding rounds dominate the loop.
• Medium exposure: SRE, security, and data engineering, where systems design still gates but a coding screen exists.
• Lower exposure: engineering management, PM, design, where behavioral and portfolio rounds dominate.

If you run reqs across all three tiers, your Q4 miss risk is not spread evenly. It is concentrated on IC engineering, and it is concentrated in the metros where onsite is now mandated. Everyone else can afford to stay remote-first through the winter. Your Staff SWE cannot.

How to rebuild a geography-first pipeline before Q4 misses

Rebuild the funnel so geography is the first filter, not the last. The playbook that worked in 2023 (national blast, filter later on interest) is exactly backwards for a market where 72.4% of loops end in a physical room.

Five moves that actually move the number:

1. Rewrite reqs with a commute-radius header. Not "Bay Area preferred." A hard "within X miles of [office ZIP]" that recruiters can pass to sourcing tools. Ambiguity here is what makes sourcers waste a week on candidates who will ghost the onsite.
2. Split your pipeline into "local now" and "relocatable within 30 days." Track them separately. The relocatable bucket needs a different message, a different close, and a housing-stipend conversation up front, not in week four.
3. Use plain-English, geography-anchored queries. Instead of Boolean strings that decay the moment LinkedIn tweaks a field, describe the person and the radius. This is where Refolk earns its keep: ask for "senior backend engineers in the East Bay who have shipped payments infra" and get a ranked list, without rebuilding the query every metro.
4. Prioritize secondary-metro engineers with relocation history. Refolk's index shows meaningful Python-engineer density outside the hubs. Candidates who have already relocated once are more likely to do it again on a reasonable package. Filter for that signal before you filter for skills overlap.
5. Front-load the onsite ask. Tell candidates in the first message that the loop ends in a room. Half your dropouts will happen there instead of at offer stage, which is a gift. The other half will self-select in.

100,069

What not to do

• Do not rely on detection alone. The best vendors flag behavior, not binaries, and Cluely's rebrand has burned the easiest open-web heuristics.
• Do not treat a signed no-AI pledge as a control. Amazon uses one, and the DOJ cases show pledges do not stop determined actors.
• Do not source Bay Area reqs from a national list and hope the geography sorts itself out. It will not, and sourcer hours are the constrained resource.
• Do not assume interviewing.io-style well-constructed problems fix this. Aline Lerner is right that good problems with probing follow-ups remain relatively cheat-resistant, but that is an interviewing argument, not a sourcing one. The onsite mandate is not going away because your problem set improved.

The Cluely and Interview Coder recruiter response question is really two questions stacked. The first is how to run a loop that verifies a human. The industry has answered that: an onsite round. The second is how to source into that loop without watching your addressable pool collapse by two-thirds. That answer is geography-first sourcing, and it is a workflow change more than a tooling change. The tools that help are the ones that let you ask for the person you want in the city you need, and skip the parts of the funnel that no longer apply.

FAQ

Is the 72.4% figure durable, or will onsite mandates soften once detection improves?

Durable through at least 2026. Detection vendors flag behavior, not the underlying binaries, and Cluely's November 2025 rebrand to a general meeting assistant means the population of cheat-tool users is now indistinguishable from legitimate note-taker users on the open web. Combine that with DOJ cases involving 300+ U.S. companies unknowingly hiring North Korean IT operatives, and the in-person round is doing security work, not just quality work. That is not a control CISOs give back easily.

How much does the addressable pool actually shrink for a Bay Area onsite Python req?

Roughly 68%. Refolk's index shows about 100,069 U.S. Python-skilled software engineers, with roughly 32% clustered in the SF Bay Area. A commute-radius filter on a Bay Area req takes the reachable pool from about 100,000 to about 32,000 before any skill or seniority overlay. NYC's proportional hit is worse (about 8% share), and secondary metros lose almost nothing because they were never sourcing national anyway.

Do coding-focused reqs suffer more than other engineering reqs?

Yes, materially. The 45% share of flagged cheating attributed to Cluely and Interview Coder is concentrated in coding rounds, where overlay quality is highest and the AI answer is closest to a passing answer. SWE, ML, and platform reqs feel the geographic collapse hardest because their loops are the ones being pulled onsite first. Engineering management, PM, and design reqs can stay remote-first longer because their loops depend on behavioral and portfolio rounds that overlays do not meaningfully help with.

What is the single highest-leverage change a sourcing team can make this quarter?

Move commute radius from a soft preference to a hard filter at the top of the funnel, and pair it with a separate "relocatable within 30 days" pipeline that gets a different message. The teams that miss Q4 will be the ones sourcing Bay Area onsite reqs from national lists and discovering the geography problem at offer stage. The teams that hit will be the ones who start every query with a ZIP code and a radius, and use plain-English sourcing to fill the two buckets in parallel.