Serving a clean markdown version of your pages, whether as .md variants or an llms-full.txt content dump, does not get you cited by AI search engines. Two controlled 2026 experiments found no meaningful citation or crawl advantage for markdown over HTML: Profound's A/B test across 381 pages on six sites saw no statistically significant difference in AI bot traffic, and a separate server-log study reported that AI search engines cited the HTML version in every case and never the .md URL. Markdown can make your content cheaper for AI agents to read, which is a real but small technical win. It is not the lever that decides citations, and it is a different question from the llms.txt index file that Google already told you to skip.
The markdown question keeps getting lumped in with the llms.txt question, and the two are not the same. This piece separates the ideas people conflate, walks through what the experiments actually measured, and tells you where a markdown variant is worth serving and where it is busywork.
Three things people call "markdown for AI," and why they differ
There are three separate proposals hiding under "markdown for AI," and only understanding the differences keeps you from wasting time on the wrong one. llms.txt is a short index of links. llms-full.txt is a full-text dump of your content. A .md page variant is a clean markdown copy of a single HTML page. None of them is a proven citation signal, but they fail for different reasons and have different legitimate uses.
We have already covered the first one. As Google's May 2026 AI optimization guide put it, you "don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities)." That verdict, and the reasons behind it, are in you don't need llms.txt to show up in AI search. This article is about the two things that guide gets confused with: serving full content or clean per-page markdown, not a curated index.
| Format | What it is | Primary use | Moves AI search citations? |
|---|---|---|---|
llms.txt | A markdown index listing your key pages with one-line descriptions | Pointing AI coding agents to the right docs | No |
llms-full.txt | Your actual page content concatenated into one large markdown file | Letting a model ingest a whole doc set in a single fetch | No, though it can reduce hallucination in direct model chats |
.md page variant | A clean markdown copy of a single HTML page, usually at page.md | Cheaper ingestion for agents that fetch it directly | No; engines cite the HTML version |
What the experiments actually found
The controlled tests run in 2026 do not support the claim that serving markdown increases AI bot traffic or citations. The most rigorous one, from the AI-visibility vendor Profound, ran a three-week A/B test across 381 pages on six websites, serving HTML to a control group and markdown to a treatment group, and tracking visits from OpenAI, Anthropic, Perplexity, Meta, and DuckAssistBot. The markdown group showed roughly one additional median bot visit per page, a difference the researchers said was not statistically significant and was driven mostly by pages that were already popular. Their conclusion was blunt: the data does not support treating format as a priority.
A separate analysis went further on the citation question specifically. Over a 14-day window it reported that HTML pages received 7.4 percent of AI bot traffic while the markdown mirror files received zero, and that no AI platform cited a .md URL, every citation pointed to the HTML page. That is a striking result, but it is a single study on a limited page set over two weeks, so treat the exact "zero" figures as directional rather than a law of nature.
Even the observation that crawlers do fetch markdown when it exists comes with a caveat. On his own site, developer Dries Buytaert logged GPTBot requesting the markdown version about 35 percent of the time while ChatGPT-User did so only 0.1 percent of the time, and he found that no AI crawler used content negotiation at all: bots only found the markdown through dedicated .md URLs. His bottom line, from one personal site's logs, was that serving markdown did not reduce crawl load (his traffic rose about 7 percent) and he could not demonstrate it improved how AI systems used his content.
What to do: Do not build a markdown-serving pipeline expecting a citation bump. The best available evidence, including one controlled A/B test, shows the effect is somewhere between tiny and none. If you already serve markdown for other reasons, keep it; if you were about to invest engineering time to chase citations, stop.
Why HTML still wins the citation, and markdown does not lose either
AI search engines cite the HTML version because that is the URL a person can actually open, and because extracting text from HTML is a problem these systems solved long ago. When ChatGPT, Perplexity, Gemini, Claude, or Grok build an answer, they retrieve pages through web search, read them, and link back to a source a user can click. A .md mirror is not the canonical page, so even a bot that reads it tends to attribute the citation to the HTML URL. This matches the single-study finding above that every citation pointed to HTML.
The flip side is that clean HTML is not a handicap. The token-heavy "noise" in HTML, navigation, scripts, styling, is trivial for a modern model to strip out. The markdown token-reduction figures that vendors cite, often an 80 percent cut on a blog post or 95 percent on a product page, are real for the systems doing the reading, but they describe inference efficiency, not retrieval or ranking. Cheaper to read is not the same as more likely to be chosen. This is the same category error schema markup makes when people claim it gets you into ChatGPT: a technical nicety gets sold as a visibility lever.
What to do: Make sure your real HTML pages are crawlable, fast, and structured with clear headings and answer-first passages. That is what determines whether an engine can extract and cite you, and it works whether or not a markdown twin exists.
What is llms-full.txt, and is it worth creating?
llms-full.txt is a single markdown file that concatenates the actual text of your key pages, not just links to them, so a model can absorb your whole content set in one fetch. It uses the same structure as llms.txt but inlines the body content instead of pointing to it. Anthropic, Stripe, and many documentation platforms publish one. The mechanics and a template are covered in what is llms.txt and do you need one.
Its genuine value is not AI search visibility. It is for direct model interactions and agents: when a developer pastes your llms-full.txt URL into a coding assistant, or a support bot needs your entire product documentation in context, one clean file beats crawling twenty HTML pages. Early adopters report it reduces hallucination in answers about their product because the model is working from authoritative source text rather than snippets. That is a documentation and agent-experience benefit, not a citation benefit, and as of July 2026 no major AI search engine treats the file as a retrieval input.
The tradeoff is maintenance. A stale llms-full.txt is worse than none, because it hands models outdated content with an authoritative veneer. For a large content site the file can also balloon past what a model will actually ingest.
What to do: Create llms-full.txt only if you run a documentation-heavy or developer-facing product and you care about how coding agents and direct chats represent you. Automate its generation from your source content so it never drifts. If you are a marketing site or local business chasing AI search citations, skip it.
Should you serve a .md version of your pages?
Serving .md variants of your pages is defensible as low-cost technical hygiene, but not as an AEO tactic. If your CMS or edge platform can emit a markdown copy of each page automatically, at effectively zero ongoing effort, there is no harm in it, and agents that prefer markdown will occasionally use it. Platforms like Cloudflare, Vercel, and WordPress have made this close to free. The honest case for it stops there.
What it will not do is raise your citation rate. Across the experiments above, the markdown variant never became the cited source, and the traffic difference was not significant. Worse, if you implement it badly, by serving markdown-only pages or hiding real content behind negotiation that crawlers ignore, you can lose the HTML page that engines actually cite. The downside of a clumsy rollout is larger than the upside of a clean one.
What to do: If markdown variants are automatic and additive, sitting alongside your HTML rather than replacing it, leave them on and move on. Do not hand-maintain .md files, do not build markdown-only routes, and do not put "serve markdown to crawlers" anywhere near the top of your AI visibility roadmap.
What actually decides whether AI cites you
Citations are decided by content quality, extractable structure, and third-party presence, not by file format. 85 percent of AI citations come from third-party sites, which is to say most of what gets you recommended is not on your server in any format. On the pages you do own, the reliable wins are the same ones AI search engines have rewarded all along: open each section with a direct, self-contained answer, use clear headings that mirror how people ask questions, and keep content fresh, since engines heavily favor pages updated in the last 30 days. The mechanics of that are in how to structure your content for AI citations.
Format is the last mile of a much longer road. A markdown file cannot make thin content authoritative, cannot manufacture the Reddit threads and review-site mentions that engines pull from, and cannot substitute for ranking well enough to be retrieved in the first place. Spend the effort there. Loudmink tracks where AI search engines pull their answers from and builds that presence across blog, Reddit, and YouTube. Plans from $99/mo.
Frequently Asked Questions
Do LLMs prefer markdown over HTML?
At inference time, models handle markdown slightly more efficiently because it carries fewer tokens, but that preference does not translate into more AI search citations. Controlled 2026 testing, including a 381-page A/B experiment, found no statistically significant difference in AI crawler traffic between markdown and HTML, and citations continued to point to HTML pages.
What is llms-full.txt?
llms-full.txt is a single markdown file that inlines the full text of a site's key pages so a model can read the whole content set in one fetch, rather than the short link index that llms.txt provides. Its real use is feeding AI coding agents and direct model chats authoritative source text, which can reduce hallucination about your product. No major AI search engine uses it as a citation signal as of July 2026.
Should I serve a markdown version of my pages?
Only if your platform generates it automatically at near-zero cost and it sits alongside your HTML. It is harmless technical hygiene, but the evidence shows it does not raise citation rates, and a botched implementation that replaces or hides your HTML can cost you the page engines actually cite.
Is markdown for AI the same as llms.txt?
No. llms.txt is a curated index of links that Google's own guidance says is not needed for AI search visibility. Serving markdown content, whether as .md page variants or llms-full.txt, is about giving models cleaner text to read, which is a separate question. Both fail as citation levers, but for different reasons.
Does serving markdown improve my clean markdown citation rate?
There is no credible evidence that it does. One log study reported that no AI platform cited a .md URL over its test window, and the strongest controlled experiment found no significant traffic gain. Treat any confident "markdown boosted citations by X percent" claim as vendor marketing until it is backed by a controlled test.
Updated for July 2026: reframed Profound as an AI-visibility vendor (now executes content, not monitoring-only).