So Outside Technology Security CMS
AnQiCMS
Online Demonstration

Only about 10% of sites have installed llms.txt, can this number indicate that it is useless?

📅 2026-11-01 👁️ 0

Only about 10% of sites have installed llms.txt, can this adoption rate number prove that it is useless? No. The adoption rate describes 'how many sites have done this', and the effect describes 'whether the probability of being cited after doing it has changed', which are two independent evidence chains. The numbers from the two studies of the same period are 10% up and down and 9.3%, both of which can only show that it has not yet become a common practice.

What does the adoption rate number indicate?

A study was conducted on nearly 300,000 domain names, with only 10.13% hosting llms.txt; another study, by top-level domain, showed that 9.3% of the first thousand domains provided the file, dropping to 8.3% among the first ten thousand, with significant differences between industries, with travel sites reaching a high of 16.0%.

This set of numbers can be concluded: First, the absolute proportion is low, and second, the distribution is extremely uneven. The low adoption rate may come from cognitive lag or deployment cost, and it has no direct relationship with whether the tool is useful. Considering 'few people use it' as 'useless' is equivalent to using popularity as a substitute for effectiveness evaluation.

How should one read when the correlation is not obvious?

Another conclusion from the same batch of data is: there is no correlation between the existence of llms.txt and the number of times it is cited by AI. Here are three points to pay attention to.

The first, the sample is observational, not a controlled experiment, the content structure, weight, and theme of the site that installed the files are all different, and the confounding factors have not been eliminated. The second, being read does not mean being cited, citation also depends on the search hit and the form of the answer, and the contribution of the reading link is easy to be invisible in statistics. The third, 'no relevance' does not mean 'no effect', nor does it mean harmful, it only means that the difference cannot be measured with the existing口径.

A more practical way to judge: treat it as a reading path

Move llms.txt to the 'Model Understanding Site Structure' layer, not the 'Improve Ranking' layer. Its function is similar to a directory for the reader: what sections does the site have, what are the key pages, and how to enter. AnQiCMS will automatically generate the site LLMs.txt, and editors do not need to manually maintain this summary; this step solves the reading path issue.

Access control belongs to another matter: which crawlers can enter and which sections are not readable, is determined by the robots rules and the site's own permissions. Discussing both in one file can create the illusion of 'installing without effect' - because the reader never gets to that step.

Cross-verify with logs and sitemap

To find out if it has been read, look for three clues:

  • Whether there is a corresponding crawling request in the server log, what is the crawling frequency and the accessed file path;
  • In the automatically generated collection range of the sitemap, whether this file is exposed together with other key pages, or blocked outside by the rules;
  • After crawling, did it extend to the main content page? Hitting a single file and reading multiple articles following the links are two completely different effects.

The site map is responsible for being 'found', llms.txt is responsible for being 'understood', and both need to be cross-verified to locate where the problem is stuck. If there are no requests in the logs, the problem is not with the file content, but with the accessibility and the path to be discovered.

Common questions

Should we install it now?If the site structure is complex and the sections are numerous, the cost is very low; if the basic inclusion has not been solved, solve the inclusion and site map first, then discuss the reading path.

Will it be considered cheating if installed?It is a document for the reader, not part of the manipulation result, the judgment standard is whether the content is consistent with the actual page.

How often is it reviewed?The site section is re-generated when there are additions or deletions, trends are observed monthly with logs, and fluctuations on a weekly basis should not be concluded accordingly.

Can it be deleted if it has no effect?It can be kept for observation first, and before deletion, confirm that there are no corresponding requests in the logs to avoid misjudgment as 'deleted is better'.

Related articles

AI answers mention the brand but do not provide a link, should the mention rate and citation rate be counted separately?

In AI answers, when mentioning a brand without providing a link, should the mention rate and citation rate be separated? Yes. Mention is the brand name appearing in the answer, and citation is the answer accompanied by a link to the source. These are different events, and merging statistics will understate or overstate visibility. This article explains the misjudgment caused by the difference in measurement criteria, provides a ledger field for separate recording, and describes two things that can be improved on the content side.

2026-11-01

Where does the difference lie in the content selection system of government portals and corporate official websites in terms of audit permissions and backup?

The emphasis of the content selection system of government portals and corporate official websites differs: the former prioritizes pre-publication review, sensitive word filtering, and permission stratification, while the latter prioritizes publication efficiency and column management. The backup range should cover data and static files, and recovery should be practiced at different rhythms. This article provides a checklist for the two types of sites.

2026-11-01

Static site generator and content system mixed, which pages are suitable for pre-generation

Static site generator compiles content into files during the build phase, with extremely low read cost, but it needs to be regenerated every time it is changed; enterprise content systems are suitable for pages that can be changed at any time and have permission. When both are used together, allocate according to the type of page: documents and landing pages are pre-generated, comments, forms, and member content remain on the dynamic side, and then use pseudo-static rules to unify the address style.

2026-11-01

Only leave security fixes and not stop supporting, when should the scheduling of the runtime environment be?

PHP version enters the stage of only providing security fixes, which does not mean ending support, but functional defects will no longer be fixed. The scheduling of runtime environments should be traced back from the end date of active support, not the final end date. This article explains the differences between the two stages, the trigger points for scheduling, and the backup and regression preparation before upgrading, as well as the different rollback methods under Docker and panel deployment.

2026-11-01

The core security version fixed over twenty issues in one go, where did the rollback patch stop on which branch?

When the core security version fixes more than twenty vulnerabilities at once, it is first necessary to clarify the statistical scope, and then to see where the patch追溯 stops on which branch. This article takes a round of repairs made public at the end of this month as an example, explaining the difference between the number of fixes and the number of announcements, the minimum support branch, and why versions below the minimum do not receive any patches, and provides the disposal sequence for sites still on the old branch and the acceptance list for self-built systems during emergency upgrades.

2026-11-01

That week, the old branch was stopped just as the new version was released, what should the site running Drupal do first?

During the week of the major version release, the old branch just reached the end of support. For sites running Drupal, what should be done first? First, accept the backup, then plan the upgrade path. The official schedule compressed the two time points into adjacent days, effectively compressing the upgrade window. This article explains why the scheduling collided, what level of backup is considered acceptable, the order in which module compatibility should be checked, and what options are left after the discontinuation.

2026-11-01

Vulnerabilities exploited in the wild must be reported within 24 hours, who should submit this report

Vulnerabilities exploited in the wild must be reported within 24 hours. This obligation extends to manufacturers and software maintainers who place products on the EU market, not all website operators. This article explains the effective time of the new regulations, early warnings, detailed announcements, and the three levels of time limits for final reports, who is considered to be the party that needs to report, and how unregulated sites can organize their emergency response procedures using the same timeline.

2026-11-01

There is an additional type of traffic from image search in the report, should this type of traffic be handled separately for content?

Is there a category of data found through image search in the report? Should the site create separate content for it? Do not open a new content line yet, but check if the image description is consistent with the text on the landing page. This article explains the statistics of this type of query initiated by images, its relationship with the original web page report, why the gray data is not aligned, and gives the judgment sequence of using a keyword library to fill in this type of query and distinguishing automatic image matching from manual image selection.

2026-11-01