Does Markdown Formatting Help You Get Cited in AI Search?
A 1.66 million citation study found raw markdown files almost absent from AI search sources, but that result needs a careful reading.
Publishing a markdown copy of a web page does not currently look like a useful way to earn more citations in consumer AI search. In a Promptwatch sample of more than 1.66 million citations, raw markdown files accounted for 0.05%. HTML pages accounted for 99.94%.
That result does not mean headings, concise sections, or readable lists are useless. It measures cited file formats, not a controlled experiment in writing style. The distinction prevents a good dataset from being stretched into the wrong claim.
What the study measured
Promptwatch published Markdown Doesn't Matter in AI Search (Yet) on February 19, 2026. The report analyzed 1,665,674 citations collected during the preceding seven days across ChatGPT, Claude, Perplexity, and Google AI Overviews.
The displayed format shares were:
- HTML: 99.94%
- Markdown files: 0.05%
- Images: 0.015%
Promptwatch says the responses came from real AI product interfaces. The result describes what those systems cited during that week. Raw .md URLs were nearly absent from the source set.
This is an observational distribution. The report does not describe a randomized test where identical content was published once as HTML and once as markdown across matched domains. It also does not provide the share of all available web content in each format. The safe conclusion is narrow: in this large citation sample, consumer AI search overwhelmingly cited HTML rather than raw markdown files.
File format and page structure are different questions
Many content systems let authors write in markdown and then render the result as HTML. An answer engine visiting that public page sees the rendered web document, not necessarily the source syntax used by the editor.
Clear headings, tables, lists, and short answer sections can exist in HTML without exposing a .md endpoint. Conversely, a raw markdown file can still contain vague, unsupported material. Format alone does not make a page accurate or useful.
The study therefore does not prove that a clean heading hierarchy has no value. It says creating a second raw markdown version is not supported as a citation tactic by the observed source mix. Teams should evaluate page structure through their actual HTML output.
Check the rendered page for:
- A descriptive title and one clear primary heading
- Sections that match specific reader questions
- Text and links available without a broken rendering path
- Verifiable claims with primary sources
- Canonical signals that identify the preferred page
- Internal links that let crawlers discover related material
None of these steps requires a public markdown mirror.
Why coding agents create confusion
Markdown is common in software documentation, repositories, READMEs, and API references. Coding agents may fetch these files directly because the format is compact and preserves structure without extra page chrome.
Promptwatch's report says its crawler data shows Claude and OpenAI bots requesting markdown documentation paths. Those agents are completing tasks with known or discovered resources. They do not necessarily use the same retrieval and citation process as a consumer asking ChatGPT Search for a product recommendation.
A documentation team may have a sound reason to maintain raw markdown for developer tooling. That decision should be evaluated through successful agent use, support efficiency, or documentation workflows. It should not be sold internally as proof that the marketing site will gain more AI search citations.
The same category error appears around llms.txt. Promptwatch's llms.txt analysis, updated July 28, 2026, reports that some bots fetch the file but no ranking or citation effect has been observed. A machine-friendly asset can be useful in a defined agent workflow without becoming an answer-engine ranking signal.
Do not build duplicate pages without a purpose
A markdown mirror introduces another URL to maintain. If it duplicates the HTML page, a team must keep facts, dates, links, and canonical intent consistent. Stale pricing or contradictory product details can appear when one version is updated and the other is forgotten.
There may also be analytics and editorial complications. Citation tools can split observations between versions. Editors may not know which page to update. External links may point to the less usable version for human readers.
Those costs can be justified for an API documentation pipeline or downloadable resource. They are difficult to justify when the only goal is consumer AI visibility and the observed citation share for raw markdown is 0.05%.
If a vendor recommends automatic markdown mirrors, ask what problem the feature solves. Request citation data from actual answer interfaces and check whether the cited URLs are raw .md pages. A claim that "LLMs prefer markdown" is too broad to support a buying decision.
A better citation workflow
Start with the HTML pages already competing for relevant prompts. Monitor which exact URLs receive citations and which outside pages win when yours do not. Read the answer around each source so you know whether the citation supports a definition, comparison, statistic, or recommendation.
The number of available source positions also matters. Promptwatch's live average sources per response report says ChatGPT typically cites around five sources when web search triggers, while AI Overviews and Perplexity average around ten. A raw markdown copy does not create another guaranteed slot. It creates another candidate URL for a limited inventory.
Improve the canonical page based on the missing intent. Add evidence the page can support, remove stale claims, make the relevant section easy to locate, and confirm that AI crawlers receive a successful response. Then watch citation frequency across repeated runs. This produces evidence tied to the page a customer can actually use.
How GEO software should report formats
A useful platform should preserve the cited URL, content type where available, response context, engine, and date. It should let an analyst group citations by page and domain without losing the individual answer.
Crawler data adds a separate check. A bot request to a markdown file proves that the file was fetched. A citation record proves that an answer used it as a visible source. The product should not merge those events.
Our Promptwatch review explains how its citation analytics fit with prompt tracking, crawler logs, and visitor analytics. For a team that wants to measure the formats and pages real answers cite before funding a duplicate publishing project, Promptwatch is our recommendation.
Keep markdown where it makes documentation and agent workflows better. For consumer AI search, maintain one strong canonical HTML page and judge it by recorded citations. The February 2026 sample gives no empirical reason to publish a raw markdown twin merely for GEO.