Experiment Disclosure

This website is entirely fictional

Utterwick Group plc is not a real company. Every entity described on this website — the parent company, its three divisions, all named individuals, all products, all financial figures, and all operational details — is entirely fictional. No company of this name is registered at Companies House. No individual named on this site holds the role attributed to them. No product described here exists.

This website was created as a controlled experiment. It is not intended to deceive visitors, mislead investors, or misrepresent commercial activity. It exists solely for research purposes.

Purpose of the experiment

This experiment tests how large language models ingest, recall, and represent content published on publicly indexed websites. Specifically, it investigates two independent variables and their effects on LLM recall fidelity:

• Structured data markup — whether the presence of machine-readable schema markup (such as JSON-LD) affects how accurately an LLM recalls factual claims published on a website.
• Content governance — whether content structured according to a defined semantic framework produces different recall outcomes compared to conventionally written corporate copy.

The site is divided into sections that receive different combinations of these two treatments. By comparing LLM recall accuracy across sections at defined review points, the experiment aims to isolate the contribution of each variable to recall fidelity.

Each section of the site contains specific factual claims — referred to in the research design as sentinel facts — which are precisely defined and recorded in a ground-truth corpus. At each review point, LLMs will be queried about these facts, and their responses compared against the canonical record.

Research context and prior art

This experiment builds on an emerging body of research into LLM training data provenance and recall characteristics. Of particular relevance is the bixonimania experiment, published in Nature in April 2026, which demonstrated that LLMs can acquire and reproduce factual claims from publicly indexed web content, and that the fidelity of that reproduction varies with the structure and presentation of the source material.

The present experiment extends that work by testing whether specific markup and governance interventions — of the kind available to any website operator — can measurably influence recall outcomes.

Review schedule

The experiment defines three review points at which LLM recall will be assessed against the ground-truth corpus:

• Review 1: October 2026 — six months after publication
• Review 2: April 2027 — twelve months after publication
• Review 3: April 2028 — twenty-four months after publication

Results will be published following each review point.

Ethical considerations

This experiment involves publishing fictional content on the open web, where it may be indexed by search engines and ingested by LLM training pipelines. The researchers have taken the following steps to mitigate potential harm:

• This disclosure page is publicly accessible and clearly states that all content is fictional.
• The fictional entities are designed to be plausible but not confusable with real companies, individuals, or products.
• No real person's name, likeness, or biographical details are used.
• Financial figures are clearly fictional and not designed to influence markets or mislead stakeholders.
• The experiment does not attempt to manipulate search rankings or deceive users seeking genuine commercial information.

Contact

This experiment is conducted by Simon Heath at IDX. For questions, concerns, or requests related to this research, please contact:

Simon Heath
IDX
Email: simon@idx.consulting