This symptom appears when a site adds llms.txt because the market talks about it, but does not build the layer that gives the file meaning.
The file can list important pages. It can guide an agent toward the right resources. It can reduce discovery noise. But it does not govern the whole corpus.
Typical signs
- the file points to a few pages but not to essential proof;
- the AI use policy is missing or generic;
- structured data does not always match visible content;
- service pages do not distinguish offers clearly enough;
- CTAs and forms remain ambiguous;
- exclusions and limits are not declared;
- historical content contradicts primary pages;
- the site has no explicit hierarchy of sources.
Risk
The site sends a signal of modernity, but not a signal of reliability. An agent can find the file and then reconstruct meaning from an incoherent corpus. The result can be a poor recommendation, wrong expertise attribution or misdirected action.
Direction of intervention
The answer is not to remove llms.txt. It is to place it inside a complete governance layer: canonical routes, proof, AI policy, structured data, internal linking, definition pages, machine files and actionable journeys.
The most direct entry point is the Lighthouse Agentic Browsing Audit when the question starts from the file. If the problem affects the broader architecture, the Agentic Readiness Audit or AI Governance and Machine Readability becomes more relevant.