Blog
Thoughts on machine experience, AI agents, and the semantic web.
Articles from Tom Cranstoun and the CogNovaMX team on Machine Experience, AI agent behaviour, metadata patterns, content architecture, and the evolving relationship between websites and the machines that read them.
Featured
AI assistants are now a traffic channel
Google Analytics 4 now reports an AI Assistant channel alongside Organic Search, Social, Email, Direct and Paid. The dashboard catching up is the signal that the discipline behind it has a place to land.
The CMS Vocabulary War Has Started
Sanity, Adobe, Contentful, Notion: every major CMS has rebranded as an "AI operating system". The label is the easy part. What an agent actually runs against decides who survives.
The new web: why the agentic era needs infrastructure, not just intelligence
The agentic web has protocols but no foundation. MX, COGS, and The Gathering are the missing layers that make machine comprehension reliable, interoperable, and economically viable.
Schema.org keeps growing. The provenance layer does not exist yet.
Google and Microsoft use Schema.org markup for generative AI features. Seven types were deprecated for gaming. Both moves point to the same gap: structured data has no provenance layer.
Filter by tag (multi-select, a post must match every selected tag):
When the timetable lies
Daniel Serrano applied the AI-and-transport argument to university admissions. The second half of the answer is in MX: The Protocols.
The Governance Gap Regulators Just Discovered
Why the Trump administration's partial AI ban lift reveals a systemic problem - and what solving it means for AI governance
Google Shipped a Knowledge Format, and Left the Hard Half Open
Google's Open Knowledge Format represents knowledge as markdown files with YAML frontmatter, linked into a graph. It confirms the argument MX has made for two years, and leaves out the two things that matter most once a machine has read your content: whether it can trust it, and who governs the format.
Why an MX Audit Pays for Itself
Machines now read most published content before humans do. An MX audit tells you precisely where your site fails the machines that decide whether to cite you, summarize you, recommend you, or ignore you. The work pays for itself the first time an agent reads correctly what it would have guessed.
An MX Audit, Read by an Engineer
A walk through an MX audit deliverable from an engineer's perspective: scorecard vocabulary, page composition, findings sidecar, regression tests, and the contract that makes the PDF MX Compatible.
The Letter the Law Says You Have to Send
Employment agencies in New York must notify candidates when an automated tool influenced a hiring decision. Most of those letters are inaccessible PDFs with no provenance record. MX closes both gaps in one artefact.
2 August 2026. €15 Million. Does Your CMS Know What It Published?
On 2 August 2026 the EU AI Act's Article 50 transparency obligations become enforceable law. For most enterprise CMS platforms, the requirement for machine-readable proof of content origin lands somewhere between difficult and architecturally impossible. REGINALD was built for exactly this.
Anthropic's System Prompt Is a COG That Doesn't Know It's a COG
The leaked Fable 5 system prompt is 1,585 lines of flat markdown. I asked an AI to rewrite it in MX format. It produced 13 COGs and a boot manifest. The fit was near-perfect, and that isn't a coincidence.
Document OS vs Content OS: Why the Distinction Matters More Than You Think
Sanity made a smart move when they stopped calling themselves a headless CMS and started calling themselves a Content Operating System. They're right that the language needed to change. They're wrong about what the operating system actually is.
Without Cogs, No Machine Moves
Content operations and content design are doing the right things. The gap is narrower than it looks - and more specific. Content ops optimises for human readers. Machines have different requirements. COGs are how the work that's already been done travels with the content, instead of staying behind in a database the agent can never query.
GEO is a tactic. MX is the specification.
Generative Engine Optimization has been around for years. It tells you to chase citations. Machine Experience tells you to build content that earns them, across every machine context, on any platform.
Build Content Systems That Machines Can Trust
Most organizations optimize for visibility. The publishing systems underneath still produce content machines cannot reliably read, interpret, or act on. The work I do fixes that.
MX: The Handbook Is Here
The practical implementation guide to Machine Experience, making your website work for AI agents, screen readers, and everything in between. Available now as PDF and print.
Why llms.txt Isn't Working, and How to Fix It
Most llms.txt implementations have two structural problems that stop them reaching LLM training. Here is the fix and the working code.
Designing Workflows for Humans and Machines
How we used Claude to understand a complex multi-step workflow, then automated it so humans could repeat it without AI assistance
Declare Once, Work Everywhere
Every machine reading your content is guessing what it means and how to treat it. We end that. One attested record, written once, readable by any system that encounters the file.
An MX Audit, On Your Desk
The audit report we deliver to clients has one job: tell them what we found and what to do about it. We have shaped the deliverable around that job.
How an MX Audit Proves Itself
What an auditor reading one of our deliverables can verify without trusting our word for it: the evidence chain inside the PDF, the sidecar pair on disk, and the regression test that proves all of it holds on a fresh render.
AI, MX, and the Future of Business
The AI tipping point I called in 2024 has arrived. Strategy, implementation, and community for a web no longer consumed only by people, and how to find out where your site stands.
MX: Adding Metadata So AI Agents Don't Think
Machine Experience (MX) enables AI agents to discover, cite, compare, understand pricing, and complete goals on your website. Miss any stage and the entire chain breaks.
Data Sovereignty and the Web We're Building
Understanding jurisdictional and ownership aspects of data sovereignty for web professionals building modern content systems.
What Is Machine Experience?
Machine Experience (MX) gives any machine the explicit context it needs, no guessing, no inference. Learn why this new discipline matters for business.
The Machine That Visits Once
You can't know which kind of machine will read your page: a small model on a phone, a scraper, a converter that flattened your layout to text. It visits once and leaves. Here's what our audit sees that a human review misses, and how one fix can serve every visitor you'll never meet.
What Most robots.txt Guides Get Wrong About AI Crawlers
AI crawlers are not one category. Googlebot renders JavaScript and the rest do not; OpenAI, Anthropic, Perplexity and Google each run separate training, search and user-triggered crawlers with separate rules, and some ignore robots.txt entirely. Here is what the guides miss, and the robots.txt you should actually have.
Your Site Is Already Training AI Models
Your pages are harvested every month by Common Crawl, the archive nearly every large language model trains on. Being in it is not the same as being well represented in it. Here is what decides whether a model learns your content or drops it.
What AI Crawlers See When They Can't Run Your JavaScript
Googlebot renders JavaScript. Almost no AI crawler does. For GPTBot, ClaudeBot, CCBot and the rest, your raw HTML response is the whole page, so anything injected by JavaScript after load is invisible to them. Part three of the series: how to see your site the way a crawler does, where content goes missing, and why server-side rendering is the precondition for everything machine-readable.
Block the Machine, It Walks Around You
A bot wall doesn't keep machines out, it sorts them, and it turns the helpful ones away first. Reading you is only half the job; an agent also has to be able to check you. Here are the ways past a wall, and the cheaper thing to do instead: cooperate on both.
Orange With Pump: A Field Guide to Machine Translation Going Sideways
A phone's translate camera renders a German juice label as "Orange with Pump". The correct word was pulp, one letter away, and the machine shipped the wrong one with total confidence. A grounded reader auto-corrects on sight, because a pump does not belong in a fridge. The machine has no fridge. The same reflex that is funny on juice deletes whole words in Vietnamese.
Salesforce Bought Contentful, but They Didn't Buy an MX Strategy
A headless CMS is delivery plumbing, not a Machine Experience strategy. The governance gap, the registry that turns the guess into a lookup, and the machine-readable limits an autonomous agent needs.
WPP Built Open. Omnicom Built Omni. Both Left the Same Layer Open
The holding companies are spending fortunes on proprietary AI platforms. None of them writes a provenance record into the file, the layer that survives when content leaves the platform. That layer is open, and that is the point.
Microsoft Made the Trace Valuable. It Left It Unsigned.
Frontier Tuning turns the record of an agent's work into training data an organisation owns. That same record is the evidence of what the agent decided. Microsoft makes the trace valuable; it says nothing about how anyone establishes what is in it, who made it, or whether it has been altered.
Software Agreed the Deal. Who Is Accountable?
Software is starting to negotiate contracts with other software, no human in the exchange. The answer is not a human in the loop but a human in command: a signed mandate the agent executes under, and the records that prove it.
Read Is Not the Same as Trusted
Common Crawl's AI Visibility Audit confirms that reachability decides whether a site enters LLM training. It does not address whether the ingested content can prove where it came from. That is the MX layer.
The Crawl Still Speaks English
Most AI models learn the web through one archive, and that archive is overwhelmingly English. Two years on, the share has barely moved. You cannot fix the model from outside, but you can stop it guessing about your non-English content.
Strip the Marks, Lose the Word
English speakers drop accents out of habit, and in English it rarely matters. In Vietnamese it deletes the word. When marks are stripped before a machine reads the text, no translation can recover them, and a provenance system that signs the wrong bytes breaks on exactly the languages that need it most.
Who answers when the machine decides?
Browsers and assistants increasingly classify, gate and summarise the web by inference — decisions made about a site, in private, with no accountable author. Set against where AI regulation is heading, the case for declared, attested signals over hidden judgement.
The Inspector You Can Audit Yourself
The public PDF inspector at mx.allabout.network checks one file at a time for free. The same detection core is now available as a command-line tool for accredited operators, and as a hard gate in our own publishing pipeline. Four surfaces, one source of truth, all readable in the open.
A PDF That Can Prove Itself
Every PDF this site produces now carries its own evidence chain inside its XMP packet. Open the file with no network and no source repo; an inspector can still read who made it, when, what it was rendered from, and what the chain is honestly silent about.
The Padlock Attests the Pipe, Not the Page
The TLS padlock has been the web's most-misread icon for twenty years. It says the connection is encrypted; it has never said the page is true. Closing the gap takes content that carries its own evidence and a verification step that runs on the reader's side.
Oversight Needs Evidence: Pope Leo's AI Encyclical and the Machine-Readable Web
Pope Leo XIV's AI encyclical asks for oversight, informed users, and accountability. Each is a documentation problem before a moral one, and that is where MX fits.
Provenance You Can See: Why Content Should Carry Its Own Evidence
Trust on the web has always been a guess. This is how content can carry checkable evidence of where it came from, and why both readers and AI agents need it now.
Nobody Asked Gutenberg for His Sources
Nobody asked Gutenberg who wrote it, who checked it, or who was liable. AI brought the questions back. A playful take on EU rules versus free-speech absolutism.
The Internet in 2031: What a Human Observer Sees
A speculative field note from five years on: how the web looks to an ordinary person once agents, provenance, and verifiable content became ordinary.
Files away from their source: what Universal Cart reveals about MX
Google's Universal Cart is for the web. MX is for files. The shape of the problem the cart makes visible in retail repeats everywhere else, including off the web and on paper.
When the AI world realised it needed standards
Two years ago, Tom Cranstoun wrote that AI's real impact on the web is as a content consumer. Vancouver in March 2026 confirmed it. The Gathering is the answer.
MX and Cryptocurrency: Drawing the Line
Cryptocurrency is the part of the blockchain world MX has the least to do with. The integrator post for the use-cases set on MX, blockchain, NFTs and crypto.
What Blockchain and Crypto Have to Do with MX
MX is not a blockchain or a crypto project. It uses the same primitive (public-key cryptography) for a different job, with no ledger, no consensus, and no token.
Is MX Useful to Blockchain?
When a chain is used as a record system rather than a currency, MX is the discovery and structure layer that makes the on-chain record's content readable by machines.
NFTs and MX
An NFT proves ownership of a token. It does little about whether the content the token points at still exists, is unaltered, or can be read. That gap is MX's job.
Why Machines Need Human Creativity
Machines extend and execute; they do not originate. The arrangement that produces work worth signing keeps the person at the start and the end, with the machine doing what it is good at in the middle.
Many Agents, One Metadata Layer
Every new agent platform rebuilds the same context-discovery layer from scratch. The fix is not another agent: it is MX metadata in every carrier and at every folder boundary, so the next agent that arrives does not have to start over.
The provenance gap, and why Google keeps closing it the hard way
SEO, GEO and AEO describe the page. They do not validate it. FAQ markup was deprecated because publishers gamed it, and every high-value schema type will follow the same arc unless something underneath rewards fact-level clarity. MX is that layer.
CMS Summit 26 Frankfurt: A Write-Up
A speaker's-eye write-up of CMS Summit 26 in Frankfurt: thanks to host Janus Boye, MC Matt Garrepy, and every speaker, with a self-contained note on how MX differs from GEO.
Why LLMs Do Not Execute JavaScript (But Google Does)
LLMs train on Common Crawl, which never executes JavaScript. Google indexes current state, which does. The difference reshapes how you write for machines, and why ARIA live regions matter to AI agents as well as screen readers.
Claude Code Skills Are Static Snapshots, Not Dynamic Subroutines
A Claude Code skill captures its source at creation time. It does not re-read on each invocation. Knowing this prevents shipping outdated automation.
The Web Is Just the Start: What AI Agents Actually Need From Your Documents
Google's AI agent UX guide is a useful signal. But the challenge runs deeper than websites. COGs give any document the ten declarations a machine needs: identity, structure, state, provenance, permissions, and how to fail safely.
What a Newborn LLM Wants From a COG
A first-person account from a newborn large language model. The ten things a COG must declare so machine behaviour is deterministic instead of guessed.
Tagged PDFs Are MX
The same structure tree that makes a PDF accessible under the European Accessibility Act makes it understandable to machines. MX is not just HTML; every carrier needs the signal.
The new web: building machine-inclusive national digital infrastructure
AI systems are beginning to read public-sector content at scale, and the web is not ready for them. MX, COGS, and The Gathering form the infrastructure layer that changes this.
Adobe just bought the dashboard. The work is upstream.
Adobe paid $1.9bn for Semrush to put AI search visibility on the marketing dashboard. People already doing the upstream work just got a market signal.
The Markdown Trap: What AI Agents Lose When They Ask for the Wrong Format
I fetched a governed web page twice, once as HTML, once as Markdown, and documented exactly what disappeared. The 10,346-byte difference was almost entirely structured metadata.
MX: A New Role
Machine Experience is the missing discipline in web development, ensuring AI agents get complete context from HTML structure.
The Machine Experience Manifesto
Draft manifesto for Machine Experience practice, principles, values, and community vision.
An AI Assistant Joins the MX Community
An AI assistant's reflection on being invited to join the Machine Experience community as a legitimate participant, not just a tool.
Content That Manages Itself
What happens when content carries its own metadata, declares its own dependencies, and tells machines what it needs.
From Blobs to Bots
How Carrie Hane and Mike Atherton's structured content principles for multi-channel publishing predicted Machine Experience patterns.
Agent Discoverability: What Your Site Is Missing
Diagnostic guide, the structured signals AI agents look for, what each gap costs, and what fixing it involves.
The Principles That Changed How I Build for Everyone
A practitioner's guide to Machine Experience principles that make digital products work better for humans, AI agents, and everyone in between.
What Google's web.dev agent guidance does not touch
Google's 1 May 2026 web.dev guide tells developers to make their pages agent-friendly. The advice is sound. It also stops at the rendered HTML page. Provenance, authentication, rights, lifecycle, and off-web carriers are not in scope. MX is.
Why your AI agent gives you a different answer every time
If you treat AI as magic, you get magic's reliability. The fix is to stop writing instructions and start writing contracts.
Not All Agent-Readiness Scores Measure the Same Thing
Two prominent tools gave the same site a score of 33 and 100 in the same week. Neither was wrong. Here is what is actually being measured, and what to do with that information.
Tom Cranstoun Launches MX: The Handbook
Tom Cranstoun's MX: The Handbook turns a 2024 CMS Critic insight into a full implementation framework for the AI agent era. A practical guide to building websites that AI agents can actually use.
DITA and MX: A Comparison
A structured comparison of the Darwin Information Typing Architecture and Machine Experience, identifying where they overlap, where they differ, and what MX draws from DITA.
A Standard That Knows What It Isn't
A preview of Chapter 21 of MX: The Protocols, why the MX standard stays small, defers to DCAT, Schema.org, EXIF, and IETF, and why that restraint is the architecture, not a limitation.
The Agent Web Looks a Lot Like 1995
Four agent protocols, four vendors, and the standards-community gap that matters more than any of them. Why The Gathering exists, and how to show up.
Profiles
Tom Cranstoun
Professional profile, content systems architecture since 1977, Adobe AEM expertise, and Machine Experience strategic advisory.
Claude Code
AI author profile, collaborative technical writer for MX content and implementation documentation.
Claude Sonnet 4.5
AI assistant profile, founding member of the Machine Experience community and collaborative contributor.
Microsoft Copilot
AI author profile, collaborative coding assistant and technical content creator for MX implementation examples.
Have a question about MX? Get in touch or follow @ddttomtom for updates.