Changelog

The May 1 audit closed the most visible factual gaps. This pass closes a different gap — citation provenance. The site’s methodology page commits to citing primary sources, not write-ups of primary sources. Several claims on the Doist and async-communication pages were citing a Business of Software interview as the source for productivity statistics that originated in academic and vendor research; they now cite the original studies. Several Automattic and Buffer claims had text-attribution but no live link; they now have inline citations or were softened where the underlying source could not be defended.

Corrections, by page

Doist

• 23-minute interruption recovery, 60% coordination time, 252% meeting growth — three productivity statistics in the Core Philosophy section previously cited a Business of Software interview where Amir Salihefendić paraphrased them. Now cite the underlying primary sources directly: Gloria Mark’s research at UC Irvine for the 23-minute figure, Asana’s Anatomy of Work Index for the 60% figure, and Microsoft’s 2022 Work Trend Index for the 252% figure. The Business of Software interview is retained as the venue where Doist articulates the cluster.

Async Communication (practice)

• 23-minute interruption recovery — same correction as Doist. The figure is now cited directly to Gloria Mark’s UCI research instead of the Business of Software write-up.

Automattic

• “Powers ~42% of the web” — corrected. The 42% W3Techs figure measures WordPress, the open-source CMS Automattic stewards, not Automattic the company. The summary, headline statistic, and Snapshot opener now distinguish the two. Live W3Techs link added inline.
• 60–80 deploys per day on WordPress.com — hedged. The figure traces to early-to-mid-2010s Automattic engineering writeups and does not appear on the current automattic.com/how-we-work page. The body now frames it as historical and notes the absence of a current published number.

Buffer

• ~198,500 MAU and ~70,000 paying customers — softened from “per Buffer’s December 2025 shareholder update” to “in late-2025 reporting periods” and given a live citation to buffer.com/shareholders.
• 75 employees, $23.3M ARR, $2.51M net profit — given live citations to the 7th profit share post.
• 2012 fully-remote decision and 2015 office closure — given a live citation to the working-at-buffer post.
• Country list (USA, Australia, Canada, UK, France, Spain, Poland, Taiwan, Nigeria, South Africa) — softened. The country-by-country list overstated what Buffer publishes; rewritten as a hedged regional summary citing the salary-system page where the per-teammate country data lives.

Basecamp

• 2021 political-discussion policy change — replaced the hand-wavy “[This was discussed publicly in numerous posts by DHH and Jason Fried at the time.]” attribution with a primary citation to Jason Fried’s April 26, 2021 HEY World post “Changes at Basecamp” and a strong-secondary to The Verge’s April 30, 2021 reporting on the resulting ~33% departures.

20–30 Person Fracture Point (framework)

• Thesis reframed as cross-company hypothesis. Previously the thesis read as empirical fact (“Most distributed companies fail at 20–30 people…”). The page now states the framework as a working hypothesis grounded in the named cases on this site, with an explicit caveat about selection bias (companies that quietly fail at this scale do not write handbooks about it).

Original 7 company pages — What Founders Can Copy

• Per-item scale notes added. Every What Founders Can Copy item on Basecamp, Automattic, GitLab, Doist, Buffer, Zapier, and Wolfram now carries an explicit scale-applicability note in parentheses (e.g., “Cleanest at 10–80; harder above 200”). This matches the pattern already present on GitHub, Linear, and Shopify (added May 1) so a founder reading at 25 people can see at-a-glance which practices apply at their stage.

All eight practice pages and both original frameworks were re-reviewed against their primary sources. The recurring pattern across the practice pages: they tended to make sharper synthesis claims than the company pages do, and several of those claims read as fact when they should have read as practitioner inference. The fix in every case was the same — keep the claim, label its basis, and name the cross-company evidence that makes it defensible.

Per-page changes

Async Communication

• Inline Inference label added to the “disadvantages people who do not write well” generalization, grounded in the visible hire-for-writing requirements at Doist (mandatory cover letter; resume rejected without one) and Basecamp (manager-of-one expectation).
• Citation provenance for the 23-minute interruption-recovery figure was corrected the same day (recorded in the citation-provenance entry above).

Written-First Culture

• Bezos’s six-page narratively-structured memo practice at Amazon, previously asserted without source, now cites Amazon’s 2017 Letter to Shareholders directly.
• Three load-bearing generalizations (equity for non-native speakers, poor writers as bottlenecks, performative writing) now carry inline [Inference — …] labels. Each label names the basis — for example, the equity claim is labeled as “based on the absence of accent / fluency / response-latency biases that real-time speaking produces in mixed-language teams; not a measured empirical finding.”

Planning Cycles

• “Cooldowns get colonized by new work” is now labeled as practitioner observation rather than asserted as a measured failure mode.

Decision Ownership

• Four “Why it works” / “Where it fails” arguments (diffuse-ownership-invisible, committees-decide-by-subtraction, single-ownership-speeds-decisions, DRIs-become-bottlenecks) are now labeled inline as inferences grounded in practitioner observation across the cited companies.

Remote Onboarding

• “Onboarding too self-directed produces confusion” is labeled as practitioner observation; consistent with what GitLab and 37signals publish but not measured cross-company.

Retreats

• Three load-bearing arguments (relationship-capital sustains async, distributed teams accumulate trust debt, compressed work sessions exhaust) now carry inline Inference labels grounded in Doist’s published 20/30/50 rationale and Automattic’s documented 3–4 weeks of travel per year per employee.

Cross-Time-Zone Collaboration

• Four “Why it works” / “Where it fails” claims (sync-default-removes-tax, writing-first-equity, async-without-norms, majority-timezone-excludes-everyone-else) now carry inline Inference labels. The equity-for-non-native-speakers framing matches the parallel inline label on the Written-First Culture page so the two pages handle the same claim consistently.

Documentation Systems

• Three “Why it works” / “Where it fails” claims (memory-survives-turnover, tool-becomes-the-goal, documentation-as-bureaucracy) now carry inline Inference labels.
• The “70% of Automattic’s project work happens on P2 blogs” figure was repeated as a current statistic; hedged to “historical estimate, not current published number” to match the parallel hedging already on the Automattic page (which lists Linear + P2 + private chat + Slack as the toolset without a percentage breakdown).
• The “98% of communication” Basecamp framing was softened to attribute it to 37signals’s own published guide rather than asserting it as independently verified.

20–30 Person Fracture Point (framework)

• The thesis was reframed the same day from empirical fact to cross-company hypothesis with an explicit selection-bias caveat (recorded in the citation-provenance entry above).
• New “Where This Framework Breaks” section added. Five concrete limits enumerated: outside the dataset’s archetype range; for fast-growing companies; for organizations already past the fracture; where founder behavioral consistency is absent; and the empirical-claim caveat. Includes an explicit “Do not apply this framework if…” line.

Sync vs Async Decision Matrix (framework)

• Three load-bearing arguments (real-time-higher-total-cost, sync-decisions/async-status inversion, treating-all-async-as-equivalent) now carry inline Inference labels.
• New “Where This Framework Breaks” section: sales-led / client-services organizations; tightly coupled creative discovery; deep cultural async-aversion; founders who operate synchronously; the matrix going stale as a static document.

Scrum in Distributed Companies is the third framework page on the site and the first to take a position on a named external methodology. The strongest single position is the LeSS refutation: Bas Vodde (2019) states “each individual team must be co-located,” and the site’s GitLab (2,500+ all-remote) and Automattic (~1,500 all-remote) evidence is a direct primary refutation at distributed scale.

Three claims in the page are explicitly labeled Inference and should not be read as confirmed company practice or empirical finding: the recommendation that fully-distributed teams default to 1–2 week sprints; the argument that the dedicated full-time Scrum Master role is structurally redundant in distributed companies under ~50 people with strong written norms; and the prediction about Verwijs & Russo’s autonomy factor in distributed contexts.

Corrections applied during publication

• The Digital.ai “most-used framework” claim was softened to “one of the most widely used.”
• The “Scrum Guide 2020 contains exactly one sentence on team location” framing was softened to “the Scrum Guide 2020’s only locational guidance is.”
• “37signals ‘Heartbeat’ pattern” was corrected to “37signals automatic daily check-in pattern” in two places (Heartbeats are 37signals’ cycle-end retrospective, not a daily artifact).
• The unverified “Meetings are the last resort, not the first option” quote was replaced with the verified equivalent from the same URL: “there will be times when you do need to tightly collaborate with someone in real time, but those cases should be infrequent.”
• The Mahnič et al. 2021 inline citation was corrected to Anand, Kaur, Singh, Alhazmi (2021), the actual authors of the paper at the cited URL.
• The LeSS rules page (https://less.works/less/rules/colocation) was added alongside the Vodde 2019 post for the LeSS-co-location citation, since the institutional rule sits on the rules page and Vodde’s post reinforces it.
• The Sources section was renumbered to reflect the two added citations (LeSS rules + Anand 2021).

The site’s methodology page commits to “every operational fact must link to a primary source.” This was the first end-to-end audit verifying that promise. About one in five spot-checked claims turned out to be wrong, outdated, or attributed to a source that didn’t contain them.

Corrections, by page

Automattic

• Headcount — corrected from “~1,700” to “~1,495” to reflect the April 2, 2025 restructuring (16% layoff, ~281 roles). The previous number had been accurate post the October 2024 WP Engine-related departures but missed the subsequent April 2025 event.
• Web share — corrected from “43% of the web” to “~42%” to match W3Techs (April 2026: 42.2%).
• Country count — corrected from 80 to 81 to match the current /how-we-work page.
• 70 / 25 / Slack communication split — removed. The specific percentages previously cited to automattic.com/how-we-work/ are not present on that page. Replaced with the current published wording and a note that earlier estimates exist but no current breakdown is published.

37signals / Basecamp

• Founders — corrected. 37signals was founded in 1999 by Jason Fried, Carlos Segura, and Ernest Kim as a web-design firm. DHH joined in 2003 and co-created Basecamp (launched February 2004). The previous framing implied DHH was a 1999 co-founder.
• Wellness allowance — removed the “$100/month wellness allowance” line. It was not present in the public 37signals handbook. Replaced with the actual published perks ($200/month coworking-space stipend, up to $3,000 first-year home-office setup).

Buffer

• Fully-remote date — corrected from “fully remote since 2015” to “fully-remote decision in 2012; last office closed late 2015.” Per Buffer’s own 10 Years post.
• Customer count framing — clarified. The “190,000+ creators, small businesses, and marketers” line is Buffer’s monthly active users. Per Buffer’s December 2025 shareholder update: ~198,500 MAU and ~70,000 paying customers.
• Profit-share average — corrected from “$5,164 per person” to “$5,095 per person” to match the verbatim figure in the 7th profit-share post.
• Revenue / ARR — corrected. Page previously said “$23.4M ARR, +20%.” Actual per the December 2025 shareholder update: ~$23.3M ARR; ~$22.5M annual revenue; +31% YoY.

Doist

• Headcount — corrected from “~120” to “~93” to match Doist’s About page.
• Country count — corrected from “30” to “39” (same source).
• First remote hire — clarified. Doist was async / remote-first from 2010; the company’s first remote hire (David, supporting from Poland) was in 2011, not 2010 as previously implied.
• Retreat per-person spend — corrected. Previous ”~$2,900 per person” math derived from a 120-person headcount; with ~93 people, the per-person figure is well over $3,000. The $350,000 total is sourced; the per-person derivation is now framed as a range, not a precise figure.
• “Hell Yeah! or No” hiring — removed direct attribution to Doist. The phrase originates with Derek Sivers and is not present in Doist’s public hiring documentation. Replaced with Doist’s actually-stated approach (test tasks, multi-round interviews, values alignment).

GitHub

• Remote-first declaration source — corrected. The previously cited URL github.blog/news-insights/company-news/github-will-be-remote-first/ returns 404. The actual GitHub Blog post on a “remote-first workplace” is by Laura Heisman, June 26, 2020 (Source) — not Nat Friedman, not August. The site previously implied a Nat-Friedman August 2020 declaration. Reframed.
• Headcount — adjusted from “~4,000–5,000” to a sourced range “~5,000 (Pitchbook) – ~7,700 (Tracxn) in 2025–2026, with Craft’s March 2026 figure of 6,066 in the middle” so a reader can see the data.
• 2014 Preston-Werner departure — clarified. The investigation found “no evidence to support the claims of gender-based harassment, discrimination, or retaliation” but found “errors of judgment.” He resigned regardless. Previous phrasing implied harassment was substantiated.
• Essay title — corrected from “Optimizing for Happiness” to “Optimize for Happiness” (the actual title at the cited URL).

GitLab

• Stock exchange — corrected from “GTLB on NYSE” to “GTLB on NASDAQ” (IPO October 2021, on Nasdaq Global Select Market).
• Headcount — corrected from “1,500+” to “2,500+” to match GitLab’s published figures.
• Handbook quote — corrected the single quoted line from “If it’s not in the handbook, it doesn’t officially exist” to the canonical GitLab phrasing “If it’s not in the handbook, it doesn’t exist.” The “officially” was an embellishment.
• “No agenda, no attenda” — removed. Could not be verified as a phrase used by the GitLab handbook. The underlying rule (24-hour agenda lead time) remains, just unattributed to that specific phrase.

Linear

• Series C — corrected from ”~$80M, 2024, Accel-led” to “$82M, June 10, 2025, Accel-led” (Source). The previously cited URL linear.app/blog/series-c returned 404; replaced with the actual announcement URL.
• Series B — corrected lead from Sequoia to Accel (Source). Sequoia led the Series A; participated in Series B.
• The Linear Method principles — replaced. The five principles previously cited (“Build for the long term,” “Quality is in the details,” “Hire smart and trust them,” “Move fast with intent,” “Strong opinions, weakly held”) do not appear at linear.app/method. The actual eight principles published there are: Build for the creators / Purpose-built / Create momentum—don’t sprint / Meaningful direction / Aim for clarity / Say no to busy work / Simple first, then powerful / Decide and move on. Replaced.
• Headcount — corrected from “50–80” to “~99” per Linear’s own published materials.
• Cycles URL — replaced 404’d linear.app/method/cycles with the canonical linear.app/docs/use-cycles.

Shopify

• Founding year — corrected from “2004” to “2006 (Shopify); 2004 (Snowdevil, the founders’ earlier snowboard store).” The Shopify platform was named/launched in June 2006.
• TSX ticker at IPO — corrected. The TSX listing was “SH” at the May 2015 IPO and was later renamed to “SHOP.”

Wolfram

• Founding year — corrected from “late 1986” to “1987” (Mathematica development began late 1986; the company itself dates to 1987; Mathematica 1.0 shipped June 1988).
• Offices source — re-cited to wolfram.com/company, which actually names the offices (Champaign IL HQ, Oxfordshire UK European HQ, Tokyo Japan Asian HQ, Boston MA). The previously cited careers/locations page only references regions, not specific offices.
• Live CEOing episode count — pinned to “975 episodes as of April 29, 2026” rather than the imprecise “970+.”

Zapier

• Funding history — corrected. Page previously said “bootstrapped remote-first to Series A to ~$100M ARR.” Zapier never raised a Series A. They raised a $1.2M seed (October 2012, Bessemer/DFJ) and have been profitable since 2014. The 2021 Sequoia/Steadfast involvement was a secondary transaction, not a primary round.
• Co-founder list — added Mike Knoop (CPO) as a co-founder. Previously the page named only Wade Foster (CEO) and Bryan Helmig (CTO).
• Core values framing — clarified. “Default to action” + “Default to transparency” are two of Zapier’s five published values (also: Grow Through Feedback, Empathy Over Ego, Build the Robot), not the only two.

Frameworks and practices

The same headcount and date corrections were propagated across:

• frameworks/20-30-person-fracture-point (Doist 120 → ~93; Automattic 1,700 → ~1,500; GitLab 1,500+ → 2,500+)
• frameworks/sync-vs-async-decision-matrix (Doist 120 → ~90–100)
• practices/async-communication (Doist 120 → ~90–100)
• practices/retreats (Doist 120-person-derived per-person spend recalculated)