Machine legibility: one discipline, four readers
The web grew a second audience that reads structure instead of pixels and greps instead of browses. Here is the thesis, and the four claims it rests on.
One discipline — make meaning machine-legible at every layer — consumed by four different readers: the indexer (ranks), the extractor (quotes), the synthesizer (associates), and the operator (acts).
Where this came from
Everything here is predicated on one observation: the Playwright MCP is one of the most widely used website-context tools in the agent ecosystem. It is how coding agents, QA loops, browser copilots, and research agents actually read pages. And what it hands the model is not the rendered page — it is the accessibility snapshot: a tree of roles, accessible names, link text, URLs, and states.
That means a de-facto standard for "what agents see" already exists, it is already deployed at scale, and it is the accessibility tree. Every claim below follows from taking that snapshot format seriously as the spec for the agent-readable web.
I noticed it the boring way — reviewing my own .playwright-mcp/*.yml
captures during a frontend build and realising I was looking at the lens every
MCP-driven agent reads the web through.
The thesis
The web is growing a second audience that reads structure instead of pixels and greps instead of browses — and everything that makes a site verifiable by agent loops (semantic landmarks, honest accessible names, provenance in payloads, machine-checkable states) is the same work that makes it legible to everyone else's agents.
That's the arbitrage window.
The four claims
Each of these stands on its own, with evidence drawn from real work rather than hypotheticals.
| Claim | Read |
|---|---|
| The accessibility snapshot is the agent-web's SERP; semantic HTML is now a ranking factor with teeth | The accessibility tree is the agent-web's SERP |
| SEO/AEO/GEO aren't different disciplines — one discipline, four consumers | One discipline, four consumers |
| Retrieval shifted to lexical, live, self-correcting agent loops — exact-match keywords are back | Grep beats embeddings |
| Five under-priced angles the acronym debate misses | Blind spots |
Why the framing matters commercially
The acronym war — SEO vs AEO vs GEO — is a fight over which consumer you are optimizing for. It obscures that roughly 70% of the work is shared. Make meaning machine-legible at every layer, and all four readers benefit; the arguments are about the remaining 30%.
That means machine legibility sells as a single service line, not four trends with four price tags. And the fourth reader — the operator, the agent that acts rather than ranks or quotes — has no acronym yet and almost no competition.
Audit a page for it
The operable half of this is already measurable. Landmarks, unnamed interactives, snapshot density, and a pass/fail verdict are all things a script can assert, which means they can fail a build — and anything that can fail a build is a deliverable rather than an opinion.
The rest of this series is the argument, one claim at a time.