TaylorMade

Writing

The accessibility tree is the agent-web's SERP

An operating agent's entire perception of a page is its accessibility snapshot. Whatever is not named there does not exist to the agent — which makes semantic HTML a ranking factor with teeth.

3 min read

Definition: An operating agent's entire perception of a page is its accessibility snapshot — roles, accessible names, link text, URLs, states. Whatever is not named there does not exist to the agent.

Predicate: the Playwright MCP — one of the most widely used website-context tools in the agent ecosystem — hands every agent this snapshot, for every page. The format is already the de-facto spec for what agents see.

The physics invert

What you're looking at in one of those snapshots is the agent-web's SERP. It isn't a rendering artifact — it's the entire universe an operating agent perceives. And its physics invert several human-web assumptions.

Everything visual — colour, hierarchy, animation, your carefully chosen condensed-italic wordmark — is gone. I proved this the hard way: the MCP literally cannot see animationName. You can ship a beautiful motion system and an agent will report a static page, correctly.

Everything named survives: roles, accessible names, link text, URLs, states.

And critically, things invisible to humans are fully visible here. I once had a test scenario match the string "Ran 2 actionsRunning" because an sr-only status label concatenated straight into the trigger's accessible name. Screen-reader text is first-class content to agents.

So the discipline that wins isn't "SEO for agents". It's the thing accessibility people have been asking for since 2001: semantic HTML is now a ranking factor with teeth.

Read your own snapshot and it shows you immediately. complementary, main, button "New session" all carry meaning. Every generic [ref=…] line is a div that spent a token to say nothing.

What this changes in practice

Three things follow, and all three are checkable rather than aesthetic.

Landmarks and accessible names are deliverables

Not "nice to have" — deliverables, because a script can fail a build on them. Landmark structure plus honest accessible names is the agent index. If a build can go red when an interactive element has no name, then naming is engineering work with a definition of done.

Kill generic wherever a semantic element exists

Generic wrapper nodes are the agent-era equivalent of render-blocking script. They cost tokens and return no meaning. Noise percentage and lines-per-interactive make the regressions visible; a ul/li with one well-named link beats three nested divs with a perfect title.

Review new UI by reading its snapshot, not only its pixels

This is the habit change that matters most, and it's the cheapest. Drive the page, read the tree, and ask whether the structure alone tells you what the page does. If it doesn't, no amount of visual polish will help the half of your audience that never sees pixels.

The uncomfortable corollary

If screen-reader-only text is first-class content to agents, then content that only agents can see is the new cloaking. That cuts against the obvious temptation to ship a shadow layer of machine-friendly markup.

The durable play is one honest semantic layer that serves both audiences. I expand on that, and four other under-priced angles, in Blind spots.

For how this claim fits the larger picture, start with Machine legibility: one discipline, four readers.

Keep reading