Do machine-readable files published for AI systems enter the reconstructions they produce? Pre-registration of a before–after study on a single entity

Server logs establish that AI systems download the declarative files a website publishes for them, llms.txt, manifests under /.well-known/, structured identity and claim files. They do not establish what happens next: a file may be downloaded without being selected by retrieval, selected without entering the context window, and present in the context without surfacing in the answer a user receives. This note pre-registers an observational study of that last link, conducted on a single entity (SEMANTIKIA) across two measurement windows: before its website and declarative files are published, and after. Responses are collected in the user interfaces of ChatGPT and Gemini, in private browsing and without conversational history, at four draws per question in each system. The question bank is fixed before the first collection and comprises 75 main questions, 10 instrument questions and 5 control questions, the latter measuring the systems’ own drift over the interval between the two windows. The instrument that tells apart a declarative file’s contribution from that of the HTML pages carrying the same statements is described in a separate annex, deposited under embargo and cited by its own DOI. This note publishes neither the question bank, nor the coding rules, nor the thresholds, nor the weighting. The study covers one entity, two systems and two windows. It can establish feasibility and an effect on this case; it cannot state a generalisable rate. The author is both experimenter and subject, a limitation the study declares rather than corrects. The note is deposited before any collection, so that the hypotheses are on record as preceding the observations. A French-language version of this note is included as a second file in this record.