What an Agent Should Do With Markdown and llms.txt in AI Search
Promptwatch data shows markdown and llms.txt files shaping how AI search reads and cites sites. The agent move is to publish clean markdown and let the crawlers read it.
AI search engines read your site differently from a human visitor. They parse the HTML, but they also look for clean markdown and for an llms.txt file that describes what you offer in plain text. Promptwatch's data on how AI search reads and cites sites shows markdown and llms.txt shaping which pages get cited. For anyone who runs an agent on their AI visibility, the practical question is what an agent should do automatically with the formats AI search actually reads.
The markdown in AI search report, published in 2026, covers how markdown and llms.txt show up in AI search citations. For anyone who cares about showing up in AI answers, the practical question is what an agent should do automatically with a format the engines read before they cite.
The signal
Promptwatch's report covers how markdown and llms.txt files shape what AI search reads and cites. The report tracks how often AI search engines read llms.txt files, how markdown pages show up in citations compared to HTML pages, and what a clean markdown page does to the chance a page gets cited.
The methodology is the part that decides what the number means. The report tracks what AI search reads and cites, not what it ranks. A page that gets read is a page the engine has in its context. A page that gets cited is a page the engine chose to surface. The two are different. A page that gets read but not cited is a page the engine has but does not use. A page that gets cited is a page the engine uses.
What the sample cannot show is the cause. A markdown page that gets cited more than an HTML page is consistent with markdown being easier for the engine to parse, not with markdown being ranked higher. The two are different. A page that is easier to parse is a page the engine understands better. A page that is ranked higher is a page the engine prefers. The report shows the read and cite pattern, not the ranking preference.
What an agent should do automatically
The first move is to publish clean markdown for the pages that matter and let the crawlers read it. An agent that manages content should keep a markdown version of your key pages, not only the HTML version. A page that is easy to parse is a page the engine understands, and a page the engine understands is a page that gets cited. An agent should flag pages that are only HTML as a gap to fix.
The second move is to publish an llms.txt file that describes what you offer in plain text. An agent that manages the site should keep the llms.txt file current, with the products, the pricing, and the facts an AI search engine needs to cite you correctly. A stale llms.txt file is worse than none, because it gives the engine facts that are wrong. An agent should flag an llms.txt file that is older than the pages it describes.
The third move is to connect the markdown to the citation. An agent should log which pages AI search reads and which it cites, and flag the pages that get read but not cited. A page that gets read but not cited is a page the engine has but does not use, and the gap between read and cited is the gap to close. The crawl to citation path is the measurement that shows the gap.
How to do it with Promptwatch
The measurement that turns the report into a site level action is the Agent Analytics view in Promptwatch. It tracks real time AI crawler logs across ChatGPTBot, ClaudeBot, PerplexityBot, GoogleOther, and the Meta AI crawler, plus the crawl to citation path and error tracking.
An agent that runs on visibility should wire the markdown report into a Unified Action. When a page gets read but not cited, the action is to check whether the page is easy to parse, whether the llms.txt file is current, and whether the page answers the query the engine ran. That is the automated move that turns a read without a cite into a fix, not a dashboard you read and forget.
What the report does not tell you
The report is a population level view. It tells you what AI search does across the whole web. It does not tell you what AI search does for your site. For that you need your own crawl and citation data joined to your own visibility data. A page that gets cited more in markdown at the population level is not a page that gets cited more for your prompts.
The report also does not say markdown is required. A page that gets cited in HTML is a page the engine can parse. Markdown is a way to make parsing easier, not a way to game the ranking. The work that follows, watching whether your markdown pages get cited more than your HTML pages, is the work your own tracking does, not the work the population report does.
The broader pattern
The markdown pattern is one instance of a wider change in AI search. The engines read more than the HTML a human sees. They read the markdown, the llms.txt file, the structured data, and the crawl logs. Each is a signal the engine uses to decide what to cite, and each is a signal an agent can manage.
For an agent that runs on visibility, that means the content that earns a citation is not only the content a human sees. The practical response is to publish clean markdown, keep the llms.txt file current, and connect the read to the cite. The agents that do that are the ones that get cited in the format the engine reads, not only the format the human sees.
What to watch next
The report is a snapshot of how markdown and llms.txt show up in AI search. The open questions are whether markdown pages keep getting cited more, whether llms.txt files keep getting read, and whether your own pages get cited in the format the engine reads. Those are exactly the questions an Agent Analytics view answers for your own domain.
The dataset Promptwatch publishes is aggregated and non identifiable, and it is refreshed constantly. That makes it useful for spotting population level shifts. It does not replace the need to track your own brand, your own prompts, and your own crawl traffic. The two work together: the public report tells you what is happening in the field, and your own tracking tells you whether it is happening to you.