Why it is uncertain
People increasingly send an agent to read a site for them. How an agent chooses what to fetch, what to believe and what to cite is not documented by anyone who builds them, and it changes with every model release.
Method
- One pipeline, measured stage by stage: fetch, retrieval, understanding, consideration, citation, action.
- Experiments run on a real commercial domain, this one.
- Hypotheses are registered before the run. Holdout sets stay sealed.
- Methodology and scrubbed results are public.
What we measure
- Whether an agent reaches the page at all, and in which representation.
- Whether its answer about the business matches what the page says.
- Whether it cites the source, and whether it can complete an action.
Results so far
The methodology, the preregistration and scrubbed results are in the open repository.
This site is the test bed. Every page has a plain-text twin, the catalogue of projects is available as structured data, and all of it is built from the same source as the pages people read. An agent and a person get the same facts.
Next
Measure this version of the site against the previous one with the same set of agent questions.