Arguments I can defend with data, about where the web is going and what it costs to arrive late.
There is no CMS behind this page. Every essay is a file in the repository, reviewed and shipped from git, which is the same claim the writing makes.
38 essays
October 6, 2026
Built with AI5 min read
I built HALO so one model writes published copy and separate models check it. Here is how that split works, and what I require before I trust a clean result.
Read the essay →October 1, 2026
Built with AI5 min read
Shopify Canvas lets merchants build a store by talking with an AI agent. The harder work is deciding what the store should say and checking that its content is true and readable by machines.
Read the essay →September 29, 2026
Transacting with AI5 min read
Shopify opened checkout to browser-based AI agents through WebMCP. I explain what that means for sites that take payments or bookings, and how to decide which actions an agent can take.
Read the essay →September 24, 2026
Built with AI5 min read
I built HALO to keep client work moving by showing who has the next move. Here is how I use it to manage projects with AI agents and keep unfinished work from sitting on my side.
Read the essay →September 22, 2026
Readable by AI5 min read
Google's AI Overviews are reducing clicks to websites. I explain what the research found and how I measure whether AI answers name a business and quote its facts.
Read the essay →September 17, 2026
Built with AI6 min read
AI agents can produce code and hand work between themselves. I explain why the costly part is checking whether the result works, and where that check belongs.
Read the essay →September 15, 2026
Readable by AI4 min read
Cloudflare's defaults changed on September 15, 2026. Learn which crawlers can read your pages and how to measure whether AI recommends your company.
Read the essay →September 10, 2026
Built with AI4 min read
A disclosed Git flaw can let a booby-trapped project folder run code when an AI coding agent opens it. Learn what business and technology leaders should check.
Read the essay →September 8, 2026
Transacting with AI5 min read
Shopping agents can search a store's catalog, build a cart and hand a shopper to checkout in one conversation. I explain what product data needs to be readable and how to check whether AI tools can find it.
Read the essay →September 3, 2026
Transacting with AI5 min read
AI agents can work in signed-in websites. I explain what this means for login pages and forms, and how I measure AI recommendation visibility and score a site's readability for agents.
Read the essay →September 1, 2026
Built with AI4 min read
An AI model reported zero cost while billing for real. I changed how my agent system checks prices, controls runs, and reports spend.
Read the essay →August 27, 2026
Readable by AI5 min read
A study found an llms.txt file on 8.7% of the top 1,000 websites. I explain Google's mixed signals and why readable pages matter more.
Read the essay →August 25, 2026
Built with AI4 min read
How I carry lessons from one AI project into the next, so the same error does not keep costing time and money.
Read the essay →August 20, 2026
Readable by AI7 min read
Every measurement layer the web adopted got a public scoreboard before it got an agreed standard, and the scoreboard is what forced the definition. Agent readiness has audits and no scoreboard. Here are the four conditions a benchmark has to meet to deserve being believed, and the conflict of interest I am declaring up front.
Read the essay →August 18, 2026
Readable by AI7 min read
Chrome's agent audit checks the accessibility tree because that tree is how software perceives a page. The structure built for screen readers is now the interface an AI agent uses to find and operate your controls, which hands a decade-old budget argument a second business case.
Read the essay →August 13, 2026
Readable by AI7 min read
The facts a buyer wants most are the ones most often locked inside a PDF, a slide deck, or a photograph of a printed page. A site can pass every technical check and answer nothing. Structure gets you read; only text gets you quoted.
Read the essay →August 11, 2026
Readable by AI7 min read
A site can pass every machine-readability check and still never be named when a buyer asks an AI system who to hire. Conformance is measured on property you own and citation is measured on property you do not. Here are the four positions the pair puts you in, and the different bill attached to each.
Read the essay →August 6, 2026
Readable by AI8 min read
The full rubric I use: fifteen mechanical checks in five groups, a hundred points, and an evidence line recorded for every one of them. Published rather than sold, because a score nobody can reproduce is worth nothing.
Read the essay →August 4, 2026
Readable by AI8 min read
Chrome added an experimental Agentic Browsing category to Lighthouse in May 2026, grading how well software can read and operate a page, while Google Search says none of those files affect ranking. Both are true, they answer different questions, and nobody has published a comparison across companies yet.
Read the essay →July 30, 2026
Built with AI5 min read
WordPress peaked at 43.6% of all websites in 2025 and posted its first decline since W3Techs began tracking in 2011. Bob Michaels, building the web since 1994, argues the CMS era is sunsetting: websites are becoming software again, built with AI, readable by AI, and transacting with AI.
Read the essay →July 28, 2026
Built with AI10 min read
Human in the loop usually means a person is present. Human in control means a person can inspect the evidence, stop the action, and own the decision. Here is the approval map to draw before you automate: action classes, named owners, evidence packets, four decision states, and the thresholds that keep it from becoming bureaucracy.
Read the essay →July 23, 2026
Built with AI9 min read
An AI agent is a stateful business process with tools and permissions, and it should be evaluated like one. Here is the four-class test matrix (normal, edge, adversarial, forbidden), the trace record, the two scores that must stay separate, and the release gates that decide production.
Read the essay →July 21, 2026
Built with AI9 min read
A multi-model strategy is a stable task contract with portable evidence, and the vendors' own documentation explains why you need one: models retire on the provider's clock, notice windows run from two weeks to six months, and compatible APIs cover the request shape, never the behavior. Here is the dependency map and the exit test.
Read the essay →July 16, 2026
Built with AI9 min read
Teams are picking knowledge architecture off a model feature. The right way to choose between retrieval, long context, and a hybrid is six variables: volume, freshness, permissions, citations, cost, and latency. Here is the decision matrix, with the honest case for each side.
Read the essay →July 14, 2026
Built with AI9 min read
Retrieval is a component. Grounding is a contract: for consequential claims, the system cites approved evidence, states uncertainty, or declines, and logs which one happened. Here is the contract, the standard behind it, and the test that tells you whether a system honors it.
Read the essay →July 9, 2026
Readable by AI8 min read
Losing the AI comparison is a finding, and findings have a method. Here is the loop I run to move the recommendation: query on a locked list, decode why the winner wins, engineer that evidence on your domain with one consolidated message, and verify by retesting. With the measured results, honestly framed.
Read the essay →July 7, 2026
Readable by AI8 min read
The competitor list in your deck and the competitor field AI compares you against are two different lists, and only the second one shapes who gets recommended. Here is the discovery protocol: three query types, five platforms, and the frequency record that turns answers into a field.
Read the essay →July 2, 2026
Readable by AI8 min read
Topic chasing is the keyword-and-coverage playbook, and it measures inputs. The head-to-head measures who AI actually recommends. Here is the difference, the evidence for both halves, and how to keep your content program while changing its finish line.
Read the essay →June 30, 2026
Readable by AI8 min read
Success in business is more customers than your competition, and buyers increasingly ask AI who to choose. That makes the competitive head-to-head the only AI visibility metric that matters. Here is the lens, and how to run it.
Read the essay →June 25, 2026
Built with AI7 min read
AI interfaces stream text, update live, and fail in new ways, and every one of those behaviors has an accessibility contract. Here is how WCAG maps onto chat and agent interfaces, and why it belongs in architecture, not remediation.
Read the essay →June 23, 2026
Readable by AI7 min read
Your site has a second audience that never sees the design: crawlers, models, and agents. Here is the executive audit of what machines can actually read, grouped into five checks, with the traps that make markup lie.
Read the essay →June 18, 2026
Readable by AI8 min read
Every indexed URL is an evidence address that search engines and AI systems already trust. Here is the migration checklist that preserves the addresses, the meaning, and the machine-readable evidence through a rebuild.
Read the essay →June 16, 2026
Readable by AI8 min read
Before any redesign, inventory every public claim by entity, audience, source, date, and approval status. The contradictions you find are the real project. Here is the audit schema.
Read the essay →June 11, 2026
Readable by AI8 min read
An AI visibility report is not a plan. Route each finding by its failure type, six types, each with its own owner, smallest publishable correction, and acceptance test, then retest on locked queries.
Read the essay →June 9, 2026
Readable by AI8 min read
A repeatable AI visibility measurement starts with a locked query library: five query families mapped to real buyer situations, versioned wording, paraphrase cohorts, and full run metadata. Here is the design canvas.
Read the essay →June 4, 2026
Readable by AI8 min read
An AI answer can print a URL the model never retrieved. Here is the URL gate that separates asserted citations from grounding evidence, and the six states every citation should pass through.
Read the essay →June 2, 2026
Readable by AI9 min read
AI visibility is five separate measurements: discovery, narrative fidelity, citation, head-to-head choice, and recommendation. Here is what to record so the number tells you what to fix.
Read the essay →May 28, 2026
Built with AI9 min read
Agent permissions are being granted by integration scope instead of business consequence. Seven action classes need a named human owner no matter how automated execution gets: external communication, publication, financial commitment, access changes, destructive data actions, production deployment, and legal representation. Here is the list, with the bounded exceptions that keep it fast.
Read the essay →How this page works
Each essay is a typed entry in a registry and a component in the repository. The index, the sitemap, and the feed all read that one source, so publishing is a commit and nothing can drift out of sync. I build client sites the same way.