Things we've worked out, written down
Decisions we've made building and running our own products, with the reasoning and the numbers behind each one. You won't find listicles here.
Explore by discipline
Why most AI pilots never ship
MIT surveyed 300 corporate AI deployments and found 95% produced no measurable financial impact. The 5% that did share a shape most pilots never take.
Internal links are the SEO work you control
A study of 23 million internal links found pages with 40 or more inbound links from their own site earned four times the search traffic. That one is yours.
What self-hosting actually costs, and when it is worth it
37signals cut its cloud bill from $3.2 million to $1.3 million by buying servers. Figma pays AWS $300,000 a day and is right to. The difference is arithmetic.
Dark patterns cost more than they capture
Regulators screened 399 online shops and found manipulative design on 148 of them. The fines are large, and the research says the tactics do not even work.
Why we make four AI models argue before we ship
Claude Octopus runs the same problem past Codex, Gemini and a local model, then flags where they disagree. How the consensus gate works, and why it holds up.
Is AI really killing SaaS?
The SaaSpocalypse argument assumes shipping the app was the expensive part. Inside a company that buys software properly, it never was.
Why Astro is a great framework for simple content sites
Most content sites run on frameworks designed for applications, and they pay for it in load time. Here is what Astro does differently, with the field data.
Where Google Ads budget actually leaks
The search terms report is where most wasted PPC spend hides. How to read it, what match types really do to your budget, and what a lead should cost you.
Speed to lead: what CRM integration is actually for
A contact form that emails an inbox is no CRM integration. The research on response time is old, consistent and brutal, and most businesses still lose leads.
Why your analytics under-reports, and what to do about it
Your analytics under-counts in patterns you can predict. Browser storage limits, consent banners and ad blockers each remove a specific, predictable slice of your data.
Getting cited by AI search
There is now peer-reviewed research on what makes an AI assistant quote your page. The tactics that win look little like what SEO has taught for twenty years.
Site speed and revenue: what the field data shows
Lighthouse scores are a lab test. Here is what real browser data says about the link between load time and money, and which fixes move the number.
Six accessibility fixes that cover almost everything
WebAIM scans a million home pages every year. Six error types account for 96% of what they find, and none of them are hard to fix.
Email that reaches the inbox
Gmail and Yahoo set hard authentication rules for bulk senders in 2024. Plenty of businesses still have not met them and cannot work out why delivery slipped.
Checkout and form friction: what people abandon over
Seventy percent of carts get abandoned. Baymard has asked why across fifty studies, and the top reasons are mostly things you can fix this month.
Where AI earns its keep, and where it quietly costs you
Stanford measured how often language models invent legal citations. The number is high enough to change how you should wire AI into a business process.
Structured data: what it does and what it does not
Google publishes real numbers from sites that added structured data. The gains are large, the work is small, and most of the advice about it online is wrong.