The agent can't use your cash register if it never picks your store.

Questions this guide answers

• What is the difference between identity and capability on the agentic web?
• Does AEO still matter when AI agents complete tasks?
• Should I add an llms.txt file to my website?
• Why do AI agents choose one brand over another?
• What comes first, being cited by AI or being usable by agents?

The two-bet frame, and the move it skips

A widely-shared argument this month splits the agentic web into two bets. Identity is the layer that tells a model who you are. Capability is the layer that lets an agent do something on your site: search inventory, check a price, complete a booking. The memorable version is that identity is a brochure and capability is the cash register, and the industry keeps shipping the brochure by default while the cash register waits. The original argument is worth reading in full.

The frame is useful. It is also missing the step that comes before both. An agent only reaches your capability layer if it already decided to send a user your way. It only trusts your identity layer if something upstream surfaced your brand as a credible answer. Both layers assume the agent is already at your door. Neither answers the question that decides everything: how did it get there, and why you instead of a competitor?

The direction of travel is real. Cloudflare reported that in June 2026 automated requests crossed 57.3% of web traffic against 42.7% from humans, the first time machines outnumbered people. When agents handle most of the visits, task completion matters more than brand recognition, and the original argument is right about that. The question is what has to be true before an agent ever runs the task.

The precondition neither layer names

The layer that decides everything is the upstream choice: which brands an answer engine names, cites, and routes to when a buyer asks a question. That choice determines whether the other two layers ever fire. You can wire the most elegant callable tool on the web, but if ChatGPT, Gemini, and Perplexity never put you in the answer, no agent ever calls it.

This is Answer Engine Optimization, and it is not the same as being cited. An engine can mention you without recommending you, and a recommendation is what routes the user. We covered that gap in Cited but Not Recommended. Being chosen, named as the answer rather than buried as a footnote, is the first move, and on the agentic web it is still the move that everything else depends on.

A file you didn't write is not a strategy

The identity critique in the original piece deserves to be sat with. A popular identity file format is auto-generated by default on millions of sites: a file the owner did not write, that no major AI system has confirmed it reads. That is not a strategy. It is a checkbox that feels like coverage.

This is the failure mode we watch teams fall into with AI visibility in general: ship the artifact, assume the outcome, never verify. A file you did not write and cannot confirm anyone reads is indistinguishable, from the outside, from doing nothing. Treating AI visibility as a signal your team can trust, rather than a vanity number, means measuring what the engines actually do, not what a plugin defaulted on. We made that argument in full in Is AI Visibility a Vanity Metric?.

Three layers, in order

There are three layers, not two, and they run in sequence. Skip the first and the other two are infrastructure for traffic you never receive.

The first layer is a loop, not a push

Getting chosen is not a one-time task; it is a loop. Measure what the engines say, diagnose which failure mode you are in, execute the fix, and verify by re-asking the same prompts after it ships. Corroboration is usually where the leverage sits: engines increasingly look for independent sources to validate a claim before repeating it, which is why AI cites consensus, not authority. A perfect identity file on your own domain loses to a competitor the engine sees validated across sources it already trusts.

Before you ship an agent-readiness checkbox

Three moves, in priority order:

• Win the choice first. Track which answer engines name you, which cite you, and which competitors they cite instead. That gap is the highest-leverage work, because it decides whether the identity and capability layers ever get a chance to matter.
• Audit what ships in your name by default. Plugins auto-generate identity files. Find out what yours says and whether it matches how you would describe your own product. Drift on an auto-generated file is worse than no file.
• Verify before you assume. For any agent-readiness artifact you ship, define how you will know it worked before you ship it. If you cannot measure the outcome, you are hedging, not executing.

The bottom line

The capability layer is real, and it is coming; the original argument is right about the direction. But the cash register only rings if the agent walked into your store. Being chosen is still the first move, and it is the one you can measure today. You can run a free audit to see which engines name your brand and which name your competitors instead (free, no credit card).

Sources

• Search Engine Journal — The Agentic Web Is Splitting Into Two Bets: Identity and Capability
• Cloudflare Radar — web traffic trends (automated vs. human)

FAQ

Related guides