The Index Goes in the Context
The Index Goes in the Context
This is the third study in our AEO research series that started with a debunking and refuses to end on one. Study 1 found that no model reads llms.txt. Study 2 found that the file's value is real and the readers are the missing half of the protocol: one sentence of affordance in the agent's own tooling took the hardest retrieval class from zero to perfect. Which left the question every agent builder actually faces: if the affordance lives on the agent's side, what is the right way to build it?
We measured the menu. Same 300-page site as Study 2, same locked-room orphan pages, same questions, and five harness designs: no affordance, the one-sentence hint, a dedicated read-the-index tool, and the index prefetched straight into the model's context, in two flavors, a curated 30-entry slice and the full 300-entry everything-list. We registered a prediction before running: prefetch wins, because two months of our other benchmark series keeps teaching the same lesson. Hand the model what it needs. Do not hope it goes and gets it.
Every deliberate design works
First, the good news for everyone: the catastrophe from Study 2 dissolves under any deliberate design. Hint, tool, or prefetch, every model solved every orphan question. Once a harness decides that the site index exists, the hardest class of retrieval failure we could construct simply goes away.
So accuracy does not pick the winner. Economics does.
The economics pick a clear winner
On Haiku 4.5, the model we are evaluating for our own site agent, the cost per solved task told a clean story: about 69,000 input tokens with no affordance, 41,000 with the hint sentence, 13,500 with the index tool, and 9,400 with the curated index sitting in the system context. The prefetch design is more than seven times cheaper than hoping, and more than four times cheaper than the hint that so impressed us in Study 2.
The mechanism is mundane and decisive. The hint still requires the model to spend a round trip fetching the index and then carry it forward in conversation. The tool is better, but a tool result cannot ride a cached system prompt. The prefetch pays for the index once, in the cheapest slot an agent has, and the model starts every task already holding the map. Our registered prediction held: carriers beat initiative, here as everywhere else we have measured it.
Bigger is not better
The everything-index arm finally answered a question two studies had to leave open: does an oversized llms.txt hurt? With the full 300-entry index in context, accuracy did not move at all, and cost jumped to 2.7 times the curated slice. No grep loops, no confusion, no drama: models read a prefetched index once and navigate. Just a steady token tax for entries the task never needed.
That gives sites with large indexes, including ours, a two-part answer: publish the complete llms.txt for whatever readers arrive, and serve your own agent a curated slice.
What we're building
These three studies were never abstract. They are the design file for a chat agent we are considering for this site: Haiku 4.5, chosen after a registered backfill found the strongest retrieval profile and zero invented facts in 372 runs; a curated site index prefetched into context, chosen by this study; the hint sentence kept as a fallback for degraded modes. Every piece of that sentence has a measurement behind it, and every measurement is public.
That is what we think agent readiness actually looks like in mid-2026: not a checklist of files, but a site and a harness designed as two halves of one protocol, with evidence deciding the wiring.
