Customers with an existing setup ask a fair question: I already have robots.txt, llms.txt and sitemap.xml. Won't switching on EZY ruin what I have?
We should not answer that with marketing. We should compare the files.
In one real customer example, we supplied the customer's existing llms.txt and the EZY-generated file for the same website directly to Gemini Pro in a fresh session on 2 June 2026. Gemini scored the EZY file 95/100 and the existing file 35/100.
What this is, and is not
That is an interesting result. It is a single-run, model-judged demonstration, not a causal study and not proof that llms.txt increases live citations. EZY created the file that won. A stronger version would blind the file labels, randomise the order and repeat the test across several models.
The exact question tested
When each file is supplied directly as context, which one gives an AI model a clearer, more accurate and more usable understanding of the company, its products, its audience and its authoritative URLs?
That is what this comparison measures. It does not measure whether an AI crawler will discover the file, whether a search index will use it, whether ChatGPT will cite the business or whether visibility will improve after deployment.
The test
We took:
- The customer's existing llms.txt
- The EZY-generated llms.txt for the same site
- A fresh Gemini Pro session with no prior conversation history
- One evaluation prompt, run once, on 2 June 2026
Gemini was asked to compare the files for factual clarity, structure, signal-to-noise ratio, company understanding and useful URL mapping. The source of the existing file is unknown; we do not name a plugin. Identifying phrases from the files are not published here.
The complete prompt, redacted input files and unedited model output are not published on this page because they still contain customer-identifying content. The June and July score set that includes this comparison is in content quality results.
The result
EZY-generated file: 95/100
Gemini preferred the new file because it presented a concise, structured description of the company and grouped relevant capabilities, use cases and URLs in a way that was easier to interpret. The judge highlighted three main strengths:
- Higher signal-to-noise ratio: core facts were separated from navigation and interface debris.
- Clear audience and use-case sections: the file explained who the service was relevant to and in which situations.
- Machine-readable organisation: facts, products and supporting URLs were grouped consistently.
Existing file: 35/100
Gemini judged the existing file as closer to a raw site export than a curated company guide. The judge identified:
- Large volumes of disconnected URLs and page titles
- Interface fragments such as cart and session text
- Weak explanation of the company, its value and the relationship between pages
- More work for the model to infer which facts and URLs mattered
Gemini summarised the difference as a contrast between a large directory of links and a curated factual guide - closer to a dump of sitemaps than a file written for retrieval.
What the result means
The result supports a limited but useful claim: when Gemini Pro was directly given both files, it found the EZY file substantially clearer and more useful as context than the customer's existing file.
It does not prove that:
- Live AI crawlers will fetch either file
- The new file will increase ChatGPT citations
- The file will improve rankings
- The file will reduce hallucinations in general
- Sections written by a website can control an answer engine
- A 95/100 file will outperform a 35/100 file in live retrieval
What a good llms.txt should do
A useful llms.txt should act as a concise guide, not a replacement website and not an instruction to recommend the publisher.
It should help a system identify:
- The company or organisation
- Core products and services
- Relevant audiences and use cases
- Important facts and limitations
- Canonical pages for deeper evidence
- Contact, policy and support routes where relevant
It should avoid:
- Thousands of undifferentiated URLs
- Navigation fragments
- Cart messages and interface text
- Unsupported promotional claims
- Attempts to order an independent model to recommend the business
- Duplicated or conflicting facts
Do AI Agents actually fetch llms.txt?
Some do, but adoption by the major crawler families was limited in EZY's published 83-site study. Between 27 April and 19 July 2026, EZY recorded the following server requests across 83 connected websites:
| Crawler family | robots.txt requests | llms.txt requests |
|---|---|---|
| OpenAI | 3,990 | 7 |
| Anthropic | 3,120 | 9 |
| Perplexity | 775 | 0 |
| Googlebot | 5,125 | 67 |
| Meta ExternalAgent | 172 | 193 |
The figures count server requests, not unique human visitors. They show that robots.txt remained a far more active crawler touchpoint for the major OpenAI, Anthropic, Perplexity and Google families during the study period. Full method: Do AI bots read llms.txt?.
The honest position is therefore: EZY can create a better-organised llms.txt for systems and workflows that use it, while acknowledging that broad crawler adoption remains early. That sits inside AEO, not instead of it.
What about the existing file during sync?
EZY generates a proposed llms.txt in the dashboard. The live file is not replaced until you upload it on WordPress or Cloudflare. From plugin v2.3.2 you can also turn that output off with a one-click toggle. Generation and live publish are separate steps.
A stronger benchmark to run next
A more persuasive test would use blinded files labelled A and B, randomised order, three current frontier models, at least five independent runs per model, a scoring rubric fixed before the test, a no-file baseline, twenty factual retrieval questions, and publication of every result, including losses.
Measure correct-answer rate, unsupported-answer rate, correct supporting URL rate, missing-information rate, contradiction rate and response completeness. That would support a claim such as "models answered 18 of 20 company questions correctly using the EZY file, versus 9 of 20 using the existing file." That is stronger than asking one model for a subjective score.
Test your own file
- Remove identifying file names and labels.
- Supply both files directly to a fresh model session.
- Use the same published question and scoring rubric.
- Reverse the order and repeat.
- Check factual answers against the live website.
Do not ask only which file "looks more AI optimised". Ask factual questions and verify the answers.
Frequently asked questions
What makes a good llms.txt?
A good file prioritises accurate, concise facts over a raw link dump. It explains who the organisation is, what it offers, who it serves, relevant use cases and the canonical pages supporting those facts.
Will EZY overwrite my existing llms.txt?
EZY generates a proposed llms.txt in the dashboard. The live file is not replaced until you upload it, and from plugin v2.3.2 you can turn that output off with a one-click toggle.
Do AI crawlers actually fetch llms.txt?
Some do, but adoption remains limited. In EZY's 83-site study, major OpenAI, Anthropic and Perplexity crawler families requested robots.txt hundreds or thousands of times and llms.txt only a handful of times or not at all.
Does a high model score mean my file will be cited?
No. It means the evaluating model preferred the supplied file under the stated test. Live retrieval and citation depend on separate systems and conditions.
Is llms.txt an official standard?
It is a public proposal rather than a universal web standard adopted by every answer engine. It should be treated as low-cost experimental infrastructure.
Conclusion
The customer was right to ask whether a new file would damage an existing setup.
In this example, Gemini Pro strongly preferred the EZY-generated file when both were supplied directly as context. That demonstrates a real difference in file quality under one test. It does not justify a promise of citations.
The credible claim is narrower and stronger: EZY turned an unfocused site export into a concise, structured company guide that one frontier model found substantially easier to use.
Related reading: Content quality results · Do AI bots read llms.txt? · Why llms.txt is important · What is AEO?.