We shipped llms.txt on NetPageTwo on a Tuesday. Two weeks later we changed half the content. This is a Field Note about what we put in the file, what we removed, what we added, and why we think the format that is becoming dominant in 2026 is not quite the format the original proposal described.
This is not a definitive guide. It is what we have observed running the file on a real brand for 14 days. Take it as a data point, not as a rule.
The original llms.txt proposal positions the file as a hint to AI crawlers about the most important URLs on a domain. The structure is a Markdown file at /llms.txt that includes:
Our day-one version followed this exactly. The file was about 600 words. It listed our pricing, visibility, geo-101, about, contact, and blog URLs. The blockquote read “Productized SEO and AI search service. AI handles the volume work, a real human runs the strategy.”
We deployed it via the Code Snippets plugin (WordPress.com Business does not allow root-level static file serving), set the content type to text/plain, and added an X-Robots-Tag: noindex header so the file would not show up in standard search results.
Verifying that the file was reachable was the easy part. Curl returned 200 with the expected content. That was day one.
Three observations from days one through seven.
Observation one: the file was getting hit, but not by who we expected.
We logged requests to /llms.txt. In the first week, we saw roughly 40 requests from a variety of crawlers. Most were monitoring services (Uptime Robot, Pingdom, our own deploy verification). Maybe 3 of the 40 came from user agents that identified as GPTBot, PerplexityBot, or similar AI-engine crawlers. The remaining 37 were either generic crawler UAs or empty UAs.
So either AI engines were fetching the file via UA strings we did not recognize as AI, or AI engines were largely not requesting the file in the first 7 days. We could not distinguish between the two with the logging we had.
Observation two: nothing changed in our citation candidacy in the first 7 days.
We tested 8 AI queries we expected to be relevant to our brand. The before-and-after results were within statistical noise. We did not see citation lift from llms.txt alone in the first week. This is roughly what we expected, because most of the AI search advice in 2026 says llms.txt is at best a leading-edge experimental signal, not a direct ranking input.
Observation three: the file made a difference in sales conversations.
This was the unexpected finding. Twice in the first week, a prospect mentioned in a fit call that they had checked our llms.txt before booking. One of them said it was the moment they decided we were not just another SEO agency claiming AI expertise. The other said the file looked like we had read the spec, which most agencies he had talked to had not.
Sales credibility was the real signal in week one. The AI citation effect was unmeasurable.
Three changes based on the week-one observations.
Change one: shorten the blockquote summary.
Our day-one blockquote was one sentence: “Productized SEO and AI search service. AI handles the volume work, a real human runs the strategy.”
We rewrote it to be tighter and more specific: “Productized SEO and AI search service. AI does the volume, a real human runs the strategy. Get found on Google and cited inside ChatGPT, Perplexity, and Gemini. No contracts.”
The change added the engine names explicitly. The reasoning: if an AI engine does parse llms.txt to summarize the brand, the engine names being present makes the summary easier to compose.
Change two: add preferred-answer capsules for key questions.
The original spec does not include this, but we added a section at the bottom called “Frequently asked questions” with 5 Q&A pairs we wanted AI engines to lift verbatim if they used llms.txt as a hint. The Q&As covered pricing, contracts, response time, who does the work, and what happens if it does not work.
This was the bet that mattered most to us. If llms.txt becomes a hint that AI engines consult, having the answers we want them to give pre-written in the file is high return. If it does not, the cost is 200 extra words in a file no one reads. Asymmetric upside, low downside.
Change three: name the canonical URLs more clearly.
The day-one version had bare URLs. The week-two version has each URL labeled with a one-line description of what the page covers. “Pricing: https://netpagetwo.com/pricing/ for productized tier pricing, what’s included, contract terms.”
This change is purely a hedge in case any AI engine uses the labels to build a sitemap-of-brand-knowledge.
Two things we left alone that some llms.txt examples include.
We did not add the entire content of the homepage as a sub-section.
Some llms.txt examples include the brand’s full About page or pricing details as embedded Markdown inside the file. This expands the file to several thousand words. Our take: if the engine wants the full content, it can fetch the URL we linked. Inlining the content is duplicate signal, and if the inlined version drifts from the actual page (because we updated the page and forgot the llms.txt), the file becomes misleading.
We did not add llms-full.txt as a second file.
Some implementations of the spec include a second file at /llms-full.txt that contains the complete content of all linked pages. We did not ship this. The reasoning is similar to the above: maintenance debt, drift risk, and the bet that AI engines will fetch the URLs if they want the full content.
If a real signal emerges that engines preferentially use /llms-full.txt for grounding, we will revisit. For now, the cost of maintaining two files in sync is higher than the speculative benefit.
Three honest takeaways.
1. The file is a credibility signal more than a citation signal in 2026.
Most of the value we have measured comes from the file being there at all. Prospects who notice it treat its presence as a signal that we take AI search seriously. AI engines that read it do not yet appear to give it heavy weight in citation ranking.
This may change. The spec is young. The signal could compound.
2. The format matters less than the content.
We initially worried about exactly which Markdown structure to use. Two weeks in, we are confident the structure matters less than whether the actual content (the summary, the URLs, the preferred answers) is accurate, current, and useful.
3. The cost of shipping it is low. The cost of not shipping it is reputational.
Twenty minutes of setup, ongoing maintenance of about an hour per quarter to keep the URLs and content current. That is the total cost. The reputational risk of NetPageTwo (a brand that sells AI citation visibility) not having an llms.txt would have been larger than any uncertain upside.
The minimum viable version is:
That covers most of what the current llms.txt observed best-practice supports. Anything more is optional.
If you do not have an llms.txt yet and you sell anything related to AI search visibility, the file should ship this week. The 20-minute version is fine. The polish can come in a second pass.
If you want us to write and ship your llms.txt as part of a broader AI search optimization engagement, book the fit call. Productized SEO and AI search at $497 a month, month to month, no contract.
Related reading:
– GEO 101: Generative Engine Optimization Explained
– What is Answer Engine Optimization (AEO)?
– FAQPage schema is the single biggest AI citation move
– About the operator
