The_question_that_already_has_an_answer

I've written about the second-audience problem: brand activations get parsed by machines before they get seen by people, and structured data is a design surface now, not an afterthought. That still holds. Generative Engine Optimisation, share-of-voice tracking, the whole apparatus of "how do I get the AI to mention my brand" is real and I'm not reopening it here.

But that argument has started showing up under a new label these past two weeks: "Agent Experience," as though renaming the problem makes it different. It doesn't. Prompt tracking, share-of-voice dashboards, GEO-as-KPI: these are visibility metrics wearing agentic vocabulary. If that's all AX means, it's not a new chapter. It's the same one with a new acronym on the cover.

The_actual_new_thing

Google's August 2026 generative UI rollout to AI Overviews marks where the real shift is happening. Generative UI in AI Overviews no longer just decides whether to mention a source. The model constructs a structured response: cards, lists, interactive widgets assembled from underlying data, not a static block of text with a link attached. It's a schema-shaped object extracted from whatever structure exists in the source.

That's not citation. That's parsing. The model isn't asking "should I mention this," it's asking "can I turn this into a typed object I can render or act on." Different failure modes entirely. Citation failure means you get skipped in a paragraph. Parsing failure means the agent either fabricates a plausible substitute for your data or drops you, because it couldn't extract anything structured. The second is worse, and almost nobody writing about AX right now builds for it, because almost nobody writing about AX right now builds the agents doing the parsing.

I do. This is the version I can actually defend with evidence, not commentary.

What_"reliably_parseable"_costs_in_practice

Every repo carries an AGENTS.md file: the canonical, tool-agnostic handbook any agent can read and execute without a human translating intent. The rule is plain: "if a rule isn't in an AGENTS.md, it doesn't apply to you." That's a design commitment, not style. The file has to be complete and unambiguous enough that an agent with no shared history can execute it correctly on the first read. Prose optimised for citation doesn't need that property. Prose meant for a machine that can't ask a clarifying question does.

The same discipline appears in module manifests (schema-bound manifest.yaml files maintained in the repo, not a wiki an agent has to discover separately), in tool schemas (strict JSON, typed fields, no ambiguity about what a call requires versus what it defaults), in protocol payloads (structured, machine-parseable). None of that exists to be cited. It exists because an agent parsing it either gets a correct, typed answer or doesn't. No partial credit for "mostly clear."

The brief that commissioned this piece carries the same pattern: a "Machine spec" section with provable claims, evidence tier ratings, flags for independent verification. Strategy documents built to be actioned by an agent, not just read by a person. I didn't add that structure for this article. It's already how the fleet writes briefs, because a brief only a human can correctly interpret will eventually get executed wrong by whichever agent picks it up next.

The_design_target_is_different,_and_that's_the_whole_point

GEO content is optimised for a probabilistic judgment: will a model, reasoning over ambiguous natural language, choose to surface this. You improve your odds with authority signals, structure, clarity, relevance. It's a ranking problem.

Agent-actionable design is optimised for a deterministic outcome: does the agent produce the correct typed output, every time, with no interpretation step. My own rate-limiting logic fails closed on a Redis outage, denying the request rather than guessing, because an ambiguous outcome there has real cost. Don't leave a machine to interpret an edge case you could have specified. Applied to a public-facing interface, the question shifts from "is my content well-structured enough to rank" to "if an agent landed on this interface right now with a task, would it get a reliable answer, or fabricate one to fill the gap I left."

What_this_means_if_you're_building

The shift is simple to state and expensive to do. Stop asking whether your content is citable. Start asking whether an agent, arriving with no prior context and no way to ask a clarifying question, could parse your interface and complete a real task correctly, first try, every time.

Most published Agent Experience advice doesn't test for that. It's still measuring visibility: mentions, share of voice, presence in summaries. These are citation metrics, not actionability ones. Actionability doesn't show up in dashboards. It shows up as a broken parse, a fabricated field, an agent that silently guessed wrong because your interface gave it nothing better to work with. You won't see that failure unless you go looking, because by design, nobody tells you when a machine quietly failed to use what you built.

I build the agents on the other side of that interface for a living. The distinction between being cited and being usable isn't theoretical for me. It's the difference between a system that works and one that only looks like it does until an agent tries to act on it.