72.4% Went Back to In-Person. Rewrite Your Commute Radius Now.

Google reinstated at least one in-person round. Amazon now makes candidates sign a no-unauthorized-AI pledge. Gartner says 72.4% of recruiting leaders are conducting interviews in person to fight fraud. If your top-of-funnel still weights "remote-OK" as the default and treats geography as a soft preference, your pipeline is already misranked for how 2026 actually hires.

What the 72.4% number actually means for sourcing

The in-person interview mandate is a fraud response, not a return-to-office memo, and it collapses the addressable candidate pool for every role along a geographic axis your Boolean strings do not weight.

Here is what triggered it:

• Gartner found 72.4% of recruiting leaders are currently interviewing in person to combat candidates who pretend to be someone else or have a stand-in complete the interview.
• Sundar Pichai told Lex Fridman that Google will introduce at least one round of in-person interviews.
• Cisco, McKinsey, and Deloitte (UK graduate programme) have all reinstated in-person rounds in the past year.
• Amazon disqualifies candidates for using generative AI unless explicitly permitted, and now gets a signed acknowledgment up front.
• Coinbase rebuilt its interviews in three phases through late 2025 and early 2026, ending in a company-wide "AI Fluency" rubric in March 2026.

The mechanism is simple. Fabric's analysis of over 50,000 candidates found cheating adoption more than doubled from 15% in June 2025 to 35% by December 2025. Metaintro logged 19,368 interviews between July 2025 and January 2026 and flagged 38.5% overall for AI-assisted cheating, with technical roles hitting 48%. Physical presence is the cheapest detector anyone has found.

48%

For sourcers, the consequence is not philosophical. Every remote round wasted on a cheater is a slot you have to refill. Weighting commute radius upstream is cheaper than adding proctoring downstream. The mandate is a stealth cost-shifter to the top of the funnel, and the funnel has not caught up.

The "remote-first" default was never a candidate signal

Only about 0.09% of US software engineers self-brand as remote in their headline. The "remote-first" filter dominating sourcing playbooks was an employer assumption, not a candidate one, so unwinding it costs far less pipeline than most teams fear.

In Refolk's index, the numbers are stark:

Read the first row against the second. Of 556,863 US software engineers, senior, and staff engineers indexed, only 524 put "remote" in their headline. Roughly 1 in 1,062. Flipping a role from remote-OK to on-site does not delete a self-identified pool. It removes a convenience assumption sourcers projected onto candidates who never asked for it.

The second surprise is Gartner's counterintuitive follow-on: 62% of candidates say they are more likely to apply when a role requires in-person interviews. The mechanism is competitive. In a market where cheating has doubled in six months, honest candidates want the filter that removes cheaters from their competition. On-site is that filter, and applying to on-site roles is how they signal they belong on the honest side of it.

On-site is not a candidate cost. It is a competitive moat that honest candidates want you to build.

Region-string Boolean is now a bug

If your commute-radius search runs on the string "San Francisco Bay Area", you are under-counting the Bay by roughly thirty times. Metro-region strings do not match how engineers self-locate; city names do.

Look at the last two rows of the table again. "New York" returns 552 matching headlines in Refolk's index. "San Francisco Bay Area" returns 18. The Bay is not thirty times smaller than New York. Bay Area engineers list Mountain View, San Jose, Palo Alto, Sunnyvale, Daly City, Oakland, or just San Francisco. Almost none of them type the phrase "San Francisco Bay Area" into their headline, because nobody who lives there talks that way.

The fix is mechanical. For every on-site metro, replace the region name with an expanded city list. A serviceable Bay Area expansion is at minimum:

• San Francisco
• San Jose
• Mountain View
• Palo Alto
• Sunnyvale
• Menlo Park
• Redwood City
• Oakland
• Berkeley
• Daly City

Do the same for the New York metro (Brooklyn, Queens, Jersey City, Hoboken, Stamford), the Seattle metro (Bellevue, Redmond, Kirkland), and Austin (Round Rock, Cedar Park). Every metro your on-site roles sit in needs its own expansion table stored somewhere your sourcers can reuse. Refolk indexes city-level headline text so you can validate an expansion in one pass instead of guessing at Boolean coverage.

Commute radius as a first-class ranking signal

Treat commute distance as a scored feature, not a hard filter. A hard "within 25 miles" cut throws out the senior candidate who will move for the right role; a soft weight puts them behind the local match without deleting them.

A workable ranking looks like this:

1. Inside the metro city list. Full weight. These are the candidates who show up to an on-site loop without a plane ticket.
2. Within the same state or adjacent metro. Half weight. Willing commuters, weekly hybrid candidates, near-relocators.
3. Elsewhere in the country, no remote signal in headline. Quarter weight. Relocation conversation required, but the 62% "more likely to apply" stat says many will take it.
4. Elsewhere in the country, explicit "remote" in headline. Deprioritize for on-site roles. This is the ~0.09% who told you what they want.

Two consequences fall out. First, the top of your ranked list is now dense with people who can physically attend the loop that Google, Cisco, and Amazon-style pledges are pushing every other employer toward. Second, your outreach message writes itself: you are not asking someone to give up remote work, you are inviting them to a process that filters out the cheaters they are already competing against.

On-site catches what remote cannot

The value of an in-person round is not the room. It is the follow-up question a candidate cannot route through a second monitor. Joel Wolfe at HiredSupport screens with the Socket.io JavaScript library specifically because it is uncommon; lately candidates ace the quiz but cannot explain their own reasoning, which he calls the AI-tool giveaway.

That is the pattern Google and Amazon are responding to. From the reporting: candidates "use AI to create code that they cannot explain" and fold under follow-up questions. Chungin "Roy" Lee, the suspended Columbia student, built Interview Coder, used it to cheat an Amazon coding test, posted the interview to YouTube, and now sells the tool for $60 a month. Amazon's response - a signed pledge that candidates will not use gen AI unless permitted - shifts legal liability but does not stop the tool. The only reliable countermeasure is a room, a whiteboard, and a follow-up question.

That is why sourcing radius, not detection software, is the leverage point. Detection is a race against a $60/month subscription that ships updates faster than any proctoring vendor. Presence is not.

What to change in your stack this week

Rewrite the geographic layer of your sourcing before you touch anything else. Detection tools, pledge language, and AI-fluency rubrics all sit downstream of the pool you hand the hiring manager.

Concretely:

• Kill region-string searches. Replace "San Francisco Bay Area" and similar with city-list expansions per metro. Validate coverage against a title-only baseline so you can see what your old string was silently dropping.
• Score, do not filter. Move commute distance from a yes/no gate to a weighted feature. Keep the out-of-metro candidates visible, ranked lower.
• Stop treating "remote" as the default. In a pool where 1 in 1,062 engineers self-identifies as remote, the default should be metro-first with remote as an opt-in slice you surface separately.
• Rewrite outreach copy for the 62%. Lead with the in-person round as a feature, not a concession. Honest candidates read it as a signal you are filtering the loop they will sit in.
• Track cheater-refill cost. If your technical loop flag rate is anywhere near the 48% Metaintro reported, calculate the sourcer hours per refilled slot and compare it to the hours a commute-radius rewrite would cost once. The math is not close.

Refolk's index is built for exactly this rewrite: title-and-city queries return headline-level counts so you can see, before you send a single message, whether your expansion covers the metro or is still leaking candidates to a broken region string.

FAQ

Does an in-person mandate shrink my candidate pool? Less than sourcing playbooks assume. Only about 0.09% of US software engineers self-brand as remote in their headline (524 of 556,863 in Refolk's index). The rest were never signaling a remote preference; employers were assuming one. Switching to on-site removes an assumption, not a self-identified pool. Gartner also finds 62% of candidates are more likely to apply when a role requires in-person interviews, so the net effect on qualified inbound can be positive.

Why is region-string Boolean broken for the Bay Area? Because Bay Area engineers list their city, not the region. Refolk's index returns 552 headlines for "New York" against 18 for "San Francisco Bay Area" on the same title set. The gap is not population, it is naming behavior. Replace region strings with city-list expansions (San Francisco, San Jose, Mountain View, Palo Alto, Sunnyvale, Oakland, and so on) for every metro you source into.

Are AI-detection tools not enough on their own? No. Fabric's analysis shows cheating adoption more than doubled from 15% in June 2025 to 35% by December 2025, and Metaintro flagged 48% of technical interviews for AI-assisted cheating. Amazon's signed pledge shifts liability but does not stop tools like Interview Coder, which sells for $60 a month and ships updates faster than proctoring vendors. Physical presence with follow-up questions is the countermeasure the largest employers are converging on.

Which companies have already reinstated in-person rounds? Google, Cisco, McKinsey, Deloitte (UK graduate programme), and Amazon are on the record. Sundar Pichai committed to at least one in-person round on the Lex Fridman podcast. Coinbase went further, rebuilding its interviews in three phases through late 2025 and early 2026 and adding a company-wide "AI Fluency" rubric in March 2026 covering Usage, Application, and Understanding Limits.