... tracked 400 users encountering our 15 most common error states. Time spent reading error text: 0.9 seconds median Time spent reading button labels: 2.1 seconds median

... text analysis across 9,000 support tickets over 12 months. Tickets containing "just": 47% Common phrases: "I just want to...", "I just need to...", "I was just trying to..." Median ticket length when "just" appeared: 34 words Median

... segmented 7,800 accounts by notification settings and compared to engagement. Notifications on (default): 61% of accounts, median 4.1 sessions/week Notifications off: 39% of accounts, median

... search box that searched everything: documents, settings, help articles, users. It was powerful and slow median 2.4 seconds to results. We tested a version that searched only documents, the most common intent, and returned results in 180ms. A separate link said "Search everything else." Results over 4 weeks: Old search: 1,900 searches/week, 2.4s median

... made one page load in 2.1 seconds instead of 0.4 seconds. It was live for 6 days before we caught it. In those 6 days, we saw something strange. Users on that page were spending more time on it median 4m 20s versus

... logged connection speed metadata across 2.1 million sessions over 60 days. Under 5 Mbps: 24% 5 20 Mbps: 37% 20 100 Mbps: 28% Over 100 Mbps: 11% 61% of sessions were under 20 Mbps. Our median asset payload per page was 4.2 MB. On a fast connection that was invisible. On a 5 Mbps connection it was 7 seconds of loading. Our office had gigabit fiber. Our users had hotel wifi, mobile hotspots, and shared connections in buildings with ... ... many people and not enough bandwidth. We built a low-bandwidth mode that loaded a text-first version with images on demand. Median

... measured cursor dwell time on destructive action buttons across 220,000 sessions. Median hover time before clicking delete: 11.2 seconds Median

... product surveys with a single question and a 5-point scale. Median time from survey appearing to answer submitted: 1.8 seconds. Median

... tracked document open events across 400,000 sessions over 90 days. Median opens before first edit: 11 Median

... analyzed 840,000 in-product search queries and compared them to our documentation titles. Median search query: 7 words Median

... analyzed sessions under 30 seconds across 1.2 million visits. Sessions where the user opened and left within 8 seconds: 17% Median time to leave for those sessions: 6.4 seconds Users who did this and never returned: 61% Most common first page in these sessions: the main dashboard Nearly a fifth of our traffic was people opening the dashboard, looking at it for six seconds, and leaving without a click. They were not confused. They were checking. They wanted one number ...

... analyzed account creation behavior across 11,000 signups over 6 months. Signups who viewed exactly one page and left: 23% Median time from signup to last action for that group: 4m 10s Most common single page: a shared document or a public example Signups who returned within 30 days: 6% Nearly a quarter of our signups were people who wanted to see one specific thing that someone had shared with them. They did not want an account. They wanted access ...

... confirmation dialogs on 14 destructive actions. We assumed they were preventing errors. We measured across 500,000 sessions: Errors prevented by confirmation: unmeasurable Users who clicked through confirmation without reading: 94% Median time on confirmation dialog: 0.6 seconds Users who reported an accidental action despite confirmation: 3.1% of all actions The dialog was not a safety net. It was a speed bump people had learned to ignore. 94% of users were clicking "Yes" before their eyes had focused ...

... analyzed 9,300 support conversations over 8 months. Median first-response time: 4h 20m. But 68% of tickets contained a follow-up from the user that included the answer a screenshot, an error code, a version number, or a clarification that made the original question answerable. The user had solved it themselves, or had found the missing detail, while waiting. Median

... segmented our analytics by device for the first time in April. 60 days of data, roughly 22,000 sessions. Desktop (14,300 sessions): Median session length: 8m 40s Pages per session: 6.2 Trial start rate: 4.1% Most common entry: homepage Mobile (7,700 sessions): Median

... measured scroll direction across 1.3 million sessions over 45 days. Downward scroll events: 52% Upward scroll events: 48% Sessions with more up than down: 31% Median scroll reversals per session: 14 Users were not reading top to bottom. They were oscillating. They scrolled down to find something, scrolled up to check something, scrolled down again. Our pages were designed as linear documents, but users were treating them as maps they kept re-reading. The pages with the highest scroll reversal ...

... analyzed touch data across 900,000 mobile sessions. One-handed use: 81% Two-handed: 14% Unknown: 5% Median thumb reach zone: bottom 40% of screen Interactions in top 25% of screen: 9% of all taps Most of our primary actions were in the top right corner, which was the hardest place to reach with a thumb. Users were either using two hands, stretching, or not tapping those actions at all. We had been designing for a screen, not a hand ...

... analyzed password change events across 22,000 accounts over 18 months. Changes clustered on day 1 3 of the month: 47% of all changes Changes clustered on the same weekday: Tuesday and Monday, 61% combined Median days between changes for users who changed repeatedly: 31 People were not changing passwords because of a security event. They were changing them on a schedule, often the same schedule, likely because another system required it or because they had a monthly routine ...

... across 12,000 accounts over 6 months. Never opened settings: 81% Opened once: 14% Opened more than once: 5% Most-changed setting: notification frequency (2.1% of users) Second: theme (1.8%) Third: timezone (0.9%) We had built 47 settings. The median user changed zero. The product was effectively one configuration for almost everyone, and we had spent years adding knobs that 95% of users never touched. The lesson was not "remove settings." It was that defaults are the product. Every setting ...

... Annual plans: 19% of customers, 36% of revenue We offered a standard 20% discount for annual. Most people took monthly. Then we looked at what happened over time. Of the customers who started monthly, 34% eventually switched to annual. Median time to switch: 7 months. And the switch rate was heavily concentrated: customers who had been active for 6+ months switched at 51%, customers under 6 months switched at 4%. The insight: annual is not a purchase decision ...

... thing badly. The study held an agent loop fixed and swapped parts across four models of different size. Its planning scaffold, a plan re-injected every turn, added 11.6 points on the repository benchmark for the smallest model, whose median run otherwise ended after five turns, two runs in three without an edit. For the largest it trimmed cost by about 30% with a two-point drop in success, because what it cut was verification the model would otherwise keep ... ... each of three tool lists. Every run passed the same four checks. With the task-tracking tools switched back on, the model was offered them in every run and called them zero. With bash alone it finished in a median

... checkpoint with the relevant decisions, constraints and next steps. The patch runs native compaction, carries that checkpoint forward verbatim, and resumes the task. In one conversation covering two PRs: 767 tool calls, 18 completed compactions, and 90.9k median input tokens per tool call roughly 39 47% below the estimated 150 170k median

... move_played had fired once in the app's entire history. The cause was arithmetic, not art: a pool weight of 8 out of 108 with a selector ticking every 7 seconds meant a ~90 second expected wait. Our median session is 59.5 seconds. We had built twelve animations that a typical session could not reach. Fixing it surfaced four more defects. But the one that still bothers me is meta: our own detectors produced ten false alarms during ...

... label format, each citing the regulation, effective date and region. Another 30 are free tools. It is exactly the kind of site Google can decide is thin. What has helped so far: Internal links. The regulation pages had a median of 1 inbound link. After adding hub pages and cross-links, the median

... products get  +29% more upvotes . Busiest day Tuesday (875, weakest reception 142); quietest day  Sunday  (189) gets the best average   250 . And it's a  micro-team business :  51% of companies 10 people , median 10. Part 2. The founders: big names, pedigrees, and geography 1,602 founders  profiled. Name-matching surfaced major-founder profiles adjacent to launches Srinivas (Perplexity), Kothari (Notion), Perkins (Canva), Mensch (Mistral), Pei (Nothing)   flagged as likely match artifacts ... ... India 160  (India nearly matches US reception). Schools: UC, Stanford, Oxford, Cornell, Tel Aviv. Part 3. What actually works at launch Teams win:  solo 123 2 3 = 188 4+ = 288; blended multi-maker avg ~227 vs