1. Introduction
The llms.txt protocol was proposed as a complement to robots.txt, allowing website operators to provide structured content summaries specifically for consumption by large language models. The protocol's proponents argued it would improve AI systems' understanding of site content, increase citation accuracy, and give publishers greater control over how their content is represented in AI-generated responses.
This paper reports on llms.txt adoption across a sample of 21 service industry websites and contextualises the finding within the broader citation data from the research programme.
2. Methodology
Each of the 21 service websites in the comparative dataset was analysed for llms.txt implementation through two methods: direct URL check (requesting /llms.txt from each domain) and content analysis (searching page source for references to llms.txt). The results were recorded in the study's master JSON dataset as a boolean field (referencesLlmsTxt).
Separately, the broader research programme's citation analysis of 94,614 URLs cited across AI platform responses was examined for llms.txt references to establish the protocol's representation in actual AI citation behaviour.
3. Results
3.1 Zero Adoption in the Study Sample
The referencesLlmsTxt field returned FALSE for all 21 sites in the study sample. No site — whether cited or uncited, whether large-scale or minimal — had implemented the llms.txt protocol. This 0% adoption rate held across all four verticals: game development outsourcing (n=8), design-as-a-service (n=5), brand design agencies (n=4), and GEO consultancies (n=4).
| Metric | Value |
|---|---|
| Sites with llms.txt implemented | 0 / 21 (0%) |
| Sites referencing llms.txt in content | 0 / 21 (0%) |
| Estimated impact score | 2.5 / 10 |
| URLs referencing llms.txt in citation corpus | 1 / 94,614 (0.001%) |
| Implementation cost estimate | ~15 minutes |
3.2 Citation Corpus Analysis
In the broader research programme, 94,614 unique URLs were identified as cited sources across AI platform responses. Of these, only 1 URL was an llms.txt page. This 0.001% representation indicates that AI platforms are not meaningfully using llms.txt as a citation source, even in the rare cases where it exists. The protocol is neither widely implemented by publishers nor meaningfully consumed by AI systems — a dual-sided adoption failure.
3.3 Impact Assessment
The trust signal framework used in the broader research programme rated llms.txt at 2.5/10 for AI citation impact, placing it among the lowest-impact signals measured. By comparison, third-party reviews scored 8/10, press coverage scored 9/10, and schema markup scored 6/10. The low impact score is consistent with the zero-adoption observation: website operators may be rationally declining to implement a signal with minimal measurable return, even when implementation cost is trivial.
4. Discussion
4.1 The Zero-Cost Adoption Paradox
The llms.txt adoption gap represents a paradox in technology adoption theory. The protocol has near-zero implementation cost (~15 minutes of technical work), no maintenance burden, no risk of negative impact, and theoretical upside potential. Standard adoption models would predict rapid diffusion for zero-cost, zero-risk innovations. Yet adoption is 0% in this sample.
Several factors may explain this paradox. First, awareness is low: the llms.txt protocol has not achieved broad visibility among service industry practitioners. Second, perceived impact is low: even practitioners aware of the protocol may judge its 2.5/10 impact as insufficient to warrant action. Third, there is no competitive pressure: when no competitor has implemented a signal, the urgency to adopt it is reduced.
4.2 Comparison with robots.txt Adoption
The robots.txt protocol, proposed in 1994, took several years to achieve widespread adoption despite similar low implementation costs. However, robots.txt addressed a clear, immediate problem (unwanted crawling consuming server resources), while llms.txt addresses a speculative opportunity (improved AI representation). The absence of a problem-driven motivation may explain the slower adoption trajectory.
4.3 Should Service Websites Implement llms.txt?
Despite the low measured impact, the minimal implementation cost means that the expected value calculation favours implementation as a hedge. A 15-minute investment with even a 0.1% chance of meaningful AI visibility improvement has positive expected value. However, the data does not support prioritising llms.txt over higher-impact interventions such as third-party review acquisition, press coverage, or content structure optimisation. llms.txt is, at best, a low-priority hedge rather than a strategic priority.
5. Limitations
The sample of 21 service websites is small and limited to four verticals. Technology adoption in other industries (e.g., media, e-commerce, education) may differ. The 94,614-URL citation corpus may not represent all AI platform citation behaviour. The impact score of 2.5/10 is an estimate based on the study's trust signal framework and has not been independently validated.
6. Conclusions
The llms.txt protocol has achieved zero adoption in the service industry sample and 0.001% representation in cited URLs, despite near-zero implementation cost. This adoption gap is consistent with the protocol's low measured impact score (2.5/10) and represents a rational market response to low perceived returns. The finding contributes to understanding technology adoption patterns in the emerging AI optimisation landscape, where practitioners face numerous potential signals of varying impact and must allocate limited implementation resources accordingly.