The piece started from a real premise — readers increasingly arrive with an AI agent attached, and most writing still assumes a single unaided reader — but the first framing built on top of that premise didn’t survive its own logic. Splitting content into three audiences (humans, humans who verify, other people’s agents) sounded tidy and turned out to misdescribe the relationship: an agent isn’t a second reader, it’s an extension of the same one. The reader used to run out of patience; the correction was to notice it now runs out of structure instead, patience still intact but delegated.
The corrected framing — three levels of depth, one reader, differing by degree of augmentation — was directionally right but untested, and it stayed a checkpoint rather than a commitment until a research pass could answer three questions: is this actually happening, is it already solved, and how strong is the claim once it’s checked. That pass turned up real prior art (AEO/GEO, llms.txt, progressive disclosure, personal AI agent products) but none of it addressed the actual gap — how an author should structure public writing for a reader whose agent is doing part of the reading. The claim survived, narrowed from “nobody does this” to “someone else already proved the pattern, just not here.”
Main takeaways
- Augmentation and audience-fragmentation are not the same claim — collapsing them into each other breaks the argument before it starts.
- An untested reframe is a checkpoint, not a draft. Three specific questions (does this happen, is it solved, how strong is it) gated moving forward.
- Prior art narrowed the claim without killing it: progressive disclosure and AEO both solve adjacent problems, in the wrong direction, for a different reader.
What I got wrong
- The instinct on the way in was to claim novelty outright. That version doesn’t survive a search — the honest claim is narrower and holds up better for it.