Five blind spots in the agent-SEO conversation
Snapshot density, internal surfaces, the sr-only inversion, affordances as backlinks, and the fact that nobody's rank tracker sees agent traffic. Each with the one move that captures it.
Five things the mainstream SEO/AEO/GEO conversation misses. Each gets the single next move that captures it.
All five follow from the same predicate — the accessibility snapshot as the de-facto agent view of the web, which I set out in The accessibility tree is the agent-web's SERP.
1. Snapshot density is the new page weight
Most teams that get this far start optimizing names. The bigger lever is density.
Look at a real sidebar snapshot and count: roughly five lines per thread —
wrapper generics, duplicated truncated text, a timestamp. Tokens-per-unit-of-
meaning is the agent-era page weight, and structure controls it, not copy.
A ul/li with one well-named link beats three nested divs with a perfect
title.
There will be an "agent Core Web Vitals", and it will be measured in snapshot tokens per completed task.
The move: measure snapshotLines, generic-noise percentage, and
lines-per-interactive. Collapse wrapper nodes into semantic lists until those
numbers drop.
2. Internal surfaces are agent-consumed by design
It's easy to dismiss an internal tool as "just internal" and hold it to a lower bar.
But if your product involves agents at all, your internal surfaces are agent-consumed by design. I watched a headless worker read an internal sidebar to pick which conversation to review. The orchestration moat is precisely how reliably agents can drive your own surfaces — which makes a smoke test asserting accessible names a form of agent-SEO for your own product.
The move: hold internal tools to the same machine-legibility bar as public pages, and keep a committed test asserting accessible names on every surface an agent drives.
3. The sr-only inversion is the new cloaking
Screen-reader-only content is fully visible to agents and invisible to humans. That cuts both ways.
Content only agents can see is the new cloaking, and answer engines will eventually police it the way search engines policed hidden text. The temptation to ship a shadow layer of machine-friendly markup is real and it is a trap.
The move: never ship agent-only content. One honest semantic layer, serving both audiences.
4. Actions are the new backlinks
For operators, a site's currency is its affordances — what can be done, and how legibly.
llms.txt, schema.org potentialAction, MCP-style site manifests: these are
the sitemap.xml of this era. Shipping them now is cheap first-mover work.
The audit is worth running on yourself before you sell it. I ran it on my own
site and found /llms.txt returning 404, /robots.txt 404, /sitemap.xml
404, and no JSON-LD anywhere. Operable, but entirely undeclared.
The move: add llms.txt and potentialAction structured data to the site
checklist this quarter.
5. Nobody's rank tracker sees agent traffic
Agent traffic shows up in server logs, not Search Console.
If you already have event-store discipline, you can instrument this today and be measuring agent visits while competitors are still arguing about acronyms.
The move: log agent user-agents and fetch patterns, and surface an "agent visits" count. It becomes a sellable report the day a client asks.
Honest caveat: this is the one blind spot on the list with no verified evidence behind it yet. The other four come from work I can point at. This one is a bet, and I'd rather label it than dress it up.
The framing all five sit inside is Machine legibility: one discipline, four readers.