What happened
In June 2026, OpenAI published “The Shift to Agentic AI: Evidence from Codex.” Its internal data shows staff moving en masse from “talking to AI” to “handing tasks to an agent that runs on its own”:
- As of June 11, 99.8% of employees’ output tokens come from Codex — a year ago it was under 10%.
- The share of staff using Codex jumped from ~40% to 98%; even outside firms see 63% of output coming from agents.
- The fastest growth is among non-engineers — adoption in legal and HR is already near 90%.
(The figures are OpenAI’s own; and “71 hours of Codex a day” means cumulative runtime across many agents in parallel, not one person running it for 71 hours.)
Why it matters to you
The trend is clear: how people use AI is shifting from “I ask, it answers” to “I delegate, it does.” As more decisions get made by an agent — fetching, comparing, concluding — before the user ever opens your site, the whole contest moves forward: from “will a human click through” to “will the agent draw on you.”
And that splits into two things you shouldn’t conflate:
- GEO fitness: are you authoritative, trustworthy, worth citing — this decides whether the agent picks you.
- Agent-friendliness: can a machine fetch, parse, and use your content (clean structure, SSR so the body is visible, a markdown/plain-text version for agents, clear factual markup) — this decides whether the agent can reach you when it wants to.
You used to only need fitness, because a human was reading; now the one reading and deciding on their behalf is an agent, so whether a machine can reach and parse you becomes a second gate.
Should you act now
Both — in order: fitness first, friendliness right behind. Without trustworthy content, no format saves you; with the content there but an agent that can’t reach or parse it, you hand the citation away. Whether you’re worth picking, and whether you’re reachable once picked — your site has to answer both next.
