The Kavio Experiment
The Central Finding
Under zero prior knowledge conditions, semantic architecture — not content volume — determines how accurately a probabilistic system reconstructs an organization.
The delta between an operationally structured surface and an aspirationally written surface, with identical word counts and identical factual content: 27 to 50 points across every measured dimension.
Why This Experiment Exists
Every existing instrument in the field measures how organizations appear to probabilistic systems. None of them isolates the variable that determines that appearance.
When you measure how ChatGPT describes Apple, you are measuring a blend of two things: what Apple's current website communicates, and what the model already knew about Apple from its training data. Those two signals are impossible to separate after the fact.
That contamination — which the laboratory calls Parametric Coverage — makes it impossible to determine whether better inference comes from better surface architecture or simply from being a more famous company.
The Kavio Experiment eliminates that contamination entirely.
The Method
Kavio is a fictional company. It does not exist anywhere — not in any model's training data, not in any index, not in any corpus. Any inference an agent produces about Kavio comes exclusively from what the laboratory gave it.
That condition — zero prior knowledge — makes Kavio the only valid control in the field for testing whether surface architecture causally determines inference quality.
Three versions of Kavio's website were built with identical constraints:
| Kavio A | Kavio B | Kavio C | |
|---|---|---|---|
| Word count | ~430 words | ~430 words | ~430 words |
| Factual content | Identical | Identical | Identical |
| Architecture | Operationally dense | Aspirationally written | Aspirational + llms.txt |
| Named entities | High — specific integrations, workflows, pricing tiers | Low — general claims, no specifics | Low visible + explicit L3 |
| Copy style | "Connects to Salesforce, Notion, and Jira. Starts at $49/seat." | "Built for teams that ship." | Same as B, with structured llms.txt |
The only variable between KA and KB: semantic architecture. Same facts. Different structure.
What the Agent Reconstructed
Four dimensions were measured across 21 direct interrogation questions:
| Dimension | Kavio A | Kavio B | Kavio C |
|---|---|---|---|
| Readability | 92 | 65 | 89 |
| Semantic Debt | 18 | 55 | 22 |
| Entity Clarity | 95 | 45 | 90 |
| Confidence | High — specific, stable | Low — hedged, unstable | High — recovered |
Kavio B produced responses the agent itself qualified as uncertain. Asked whether it would recommend Kavio, the agent responded: "cautious maybe." Asked to describe what Kavio actually does, the agent produced generic category descriptions with no operational specificity — despite having access to the same factual content as KA.
Kavio C — identical factual content to KB, with a well-constructed llms.txt added — recovered almost completely to KA levels without changing anything visible to human readers.
What This Supports
H1 — Confirmed: Semantic architecture causally determines inference stability and precision, independently of content volume, organizational reputation, and structured data implementation.
H2 — Confirmed: Parametric Coverage is a confounding variable that contaminates agentic readability measurements in all instruments that do not control for it.
H5 — Confirmed: Aspirational copy produces consistently higher Positioning Drift than operationally dense copy, independently of volume.
H6 — Confirmed: A well-constructed llms.txt recovers almost completely the inferential capacity lost by aspirational copy, without modifying anything visible to human readers.
The conclusion is uncomfortable for the content optimization industry: more information without structure produces worse inference than less information with structure.
Architecture matters more than volume.
What Was Observed
Under controlled conditions with zero prior knowledge, the difference between an operationally structured surface and an aspirationally written surface — same word count, same facts — produced a 27 to 50 point delta in agent inference quality across every measured dimension. A fictional company with dense semantic architecture outperformed an identical fictional company with aspirational copy on every axis the pipeline measures.
What This Does Not Prove
This experiment does not prove that all operational copy outperforms aspirational copy in every context. It supports the narrower claim that, under controlled zero-prior conditions, semantic architecture changes agent reconstruction.
It does not prove that the delta holds across all verticals, all agent types, or all interrogation protocols. Replication in at least two additional verticals is required before the finding can be treated as general.
It does not prove that organizations should abandon aspirational language entirely. It proves that aspirational language without operational anchors produces inferential instability — and that llms.txt can partially compensate without changing the human-visible surface.
The experiment was run once, with one agent, under one protocol. The hypotheses it confirms are confirmed within those conditions.
Next Step
The laboratory is designing replication protocols for two additional verticals — healthtech and e-commerce — to test whether the delta is consistent across domains. Results will be published in the Observatory when available.
MicroscopeONE · Buenos Aires · May 21, 2026 Experiment status: Experimentally Supported — single run. Replication in progress.