Boloop-made explanatory illustration: a compact guide links to eligible published content while the CMS remains the source of truth.
When someone asked whether Boloop had llms.txt, the site did not yet have that entry point. I later decided to add a lightweight guide: explain who maintains the site and what it contains, then provide some context and links to public pages.
The result is a text entry point you can read directly. It did not give us a new “AI indexing channel” or establish that citations increased. The first problem it solved was organizing and maintaining the guide itself.
Give it one small job
Boloop has original engineering Skills, software recommendations, original articles and interactive work. Readers can explore through navigation; tools reading the site also need a concise way to understand how these parts relate.
I did not want to concatenate the whole site into an ever-growing file. The guide keeps only the introduction, three-language entry points, sections, and titles and summaries of eligible published pages. Detailed content, authorship and licensing remain on the linked pages. Third-party recommendations still belong to their original creators; being listed in our directory does not change that relationship.
Content-detail links mainly use English pages, while the language section links the English, Simplified Chinese and Traditional Chinese home pages separately. It is a compact entry point, not another stored copy of three complete bodies.
Let the CMS determine what can appear
A manually maintained text file is easy to start, but adds another place publication can drift: an article title changes while the guide keeps the old one, or an item is withdrawn while the file still exposes its link.
The Worker therefore generates /llms.txt from CMS public publication snapshots on request. It shares the sitemap discovery query’s basic eligibility rules: published, public, not deleted, backed by valid publication data, and not a demo record.
Titles and summaries come from the publication snapshot, so draft edits do not appear early. Even a request from the site owner cannot list private records visible to the administrator. Referral-detail pages for external articles and recommended Skills are not expanded as independent entries; section links still help readers find those recommendations.
Sharing basic rules does not give the two files the same format or purpose. The sitemap lists URLs for search discovery. This guide offers reading paths with short descriptions.
A small feature still needs withdrawal and pagination checks
One important risk in dynamic generation is a stale cache: content is withdrawn, but the file still lists it. Boloop does not reuse a shared snapshot for this entry point. It checks public eligibility each time and uses a private, no-store response policy.
The query must not stop at the first batch either. We reused ID-based batched reads, and isolated tests included more than one batch of records to check that later items still entered the result. Those synthetic records tested pagination; they were not the site’s actual content count.
Titles and summaries also need text handling. Newlines, brackets and Markdown symbols must not turn an ordinary title into a new file structure or a forged link. The generator cleans control characters and escapes structural characters.
If the database read fails, the entry point returns an error instead of presenting an old static copy as a current guide. This makes failure easier to identify and avoids mistaking “some text is still returned” for content consistency.
Release checks concerned file behavior
Release checks on October 22, 2026 confirmed UTF-8 plain text, readable GET content, no body on HEAD, and explicit responses to unsupported methods and paths. Public-site and non-public modes were checked separately too.
The live guide contained 43 links at that release, and every one returned a normal page. That count belongs to the release snapshot and changes as content is published or withdrawn. Checks also confirmed that the feature had not rewritten the existing robots file or sitemap.
The file carries noindex, follow, which does not make the guide broken. Its purpose is to be read by tools as needed; we are not treating the text file as an article that needs to appear in ordinary search results.
Withdrawal, draft isolation and the owner’s inability to bypass public eligibility were tested with isolated data. We did not arbitrarily withdraw production articles for acceptance testing. Together, those checks establish implementation and access behavior, not whether an AI platform actually uses the file.
Leave the benefits to observation
The llms.txt proposal website now describes broader guidance and Markdown-content suggestions. Boloop implemented a compact guide. It did not add a Markdown mirror for every page or claim complete coverage of the proposal’s suggestions.
Meanwhile, Google’s AI search guidance says that its AI search features require no additional special machine-readable files or AI text files. Google’s statement cannot, in turn, speak for every other AI tool.
I will keep later observations separate: can the file be accessed, did a tool actually read it, did an answer really cite the site, and was there attributable traffic? The site appearing in a branded query cannot automatically be credited to llms.txt.
For Boloop, the feature’s reasonable size is what it is now: reuse the CMS, keep public content consistent, and make a small guide easy to understand. Real usage evidence can inform whether it needs to grow. Protecting useful body content, clear internal links and normal page access is more valuable than attaching an unmeasured growth story to this entry point.