LLMs.txt for Digital Marketing: A 2026 Practitioner's Take (and the File I Deployed on This Blog)

Two weeks ago I added /llms.txt to aimarketingfocus.com. Then I sat on my hands and waited.

The honest answer is: nothing visible happened. No spike in AI citations. No jump in Perplexity or ChatGPT referrals. No traffic bump I could attribute to the file at all. And that's the most useful thing I can tell you about LLMs.txt right now, because most of what's being written about it is either breathless adoption copy or dismissive hot takes, and neither matches the reality.

LLMs.txt is a real, well-specified file format. It is also — as of mid-2026 — unconfirmed by any major AI engine as a signal they consume. Both of those things are true, and marketers need to plan around both.

What LLMs.txt actually is

Jeremy Howard (Answer.AI co-founder) proposed the spec on September 3, 2024. It's a plain Markdown file at /llms.txt that gives LLMs a curated, described short-list of your site's content — the way robots.txt gives crawlers access rules and sitemap.xml gives them URL lists.

The structure is dead simple: an H1 with the site name, a blockquote summary, then H2-keyed sections linking to your most important pages with one-sentence descriptions. An optional ## Optional section signals content the LLM can drop when its context window gets tight. A sibling file, /llms-full.txt, can carry the full Markdown of every linked page for deep retrieval — though most sites skip this and for good reason (more on that below).

Cloudflare serves one at developers.cloudflare.com/llms.txt and has product-level llms.txt files for nearly 100 sub-products. Firecrawl's docs ship one. Anthropic's site does not. That's roughly the adoption landscape: a few credible technical publishers, lots of vendor talk, no engine endorsement.

Why the file exists at all

The rationale Howard gave when he proposed it is worth taking seriously even if you decide not to deploy: LLMs don't read websites the way humans or even traditional crawlers do. A crawler sees your full HTML, your JavaScript-rendered DOM, your nav, your ads, your cookie banner. An LLM trying to summarize or cite your site via retrieval often gets a noisy mess that wastes its context window on chrome and reduces its ability to actually use your content.

llms.txt is the LLM equivalent of a "skip to the good parts" version of your site. You, the publisher, decide what counts as the good parts and write the descriptions in your own voice. That's the only thing it does. It does not crawl-control anything. It does not feed structured data. It is not a schema.org replacement.

The honest state of play

Here's the part most coverage glosses over:

• No major AI engine has confirmed reading /llms.txt. OpenAI (GPTBot, OAI-SearchBot), Anthropic (ClaudeBot), Google (Google-Extended), and Perplexity (PerplexityBot) have made no public statement about consuming it. The file is, at best, a courtesy signal to engines that might one day choose to honor it.
• No peer-reviewed case studies showing citation lift exist. Anecdotal GEO marketing posts claim big wins, but none isolate llms.txt as the causal variable. If you see a number attached to this file, assume it's vendor copy.
• The strongest organic adopter is Cloudflare — which is also the company whose CEO has been the most vocal advocate. That's a hint, not proof.
• llms-full.txt is a content-licensing decision in disguise. Publishing your full Markdown gives any scraper a clean training corpus. Some sites publish it; most should not.

So why deploy it anyway?

The argument for deploying anyway

Three reasons, none of them "AI search will reward you":

1. It's cheap. A static site generates one from your content collection in an afternoon. For this blog, I added a single build step that walks src/content/posts/ and writes the 30 most-recent posts with their summaries to /public/llms.txt. Total cost: about 40 lines of script and a CI rebuild.

2. It's defensive. Even if no engine reads it today, the file is a curated, described view of what you want an LLM to see when it lands on your domain. If a future AI system scrapes your site, the file is already there saying "here's the canonical summary, here are the canonical URLs." That's strictly better than letting the model guess from your footer.

3. It forces an editorial decision. Writing the H1, the summary blockquote, and the one-sentence descriptions per URL is, itself, a useful exercise. It's the same work you'd do for an internal linking audit or a content pillar page — choosing what's worth surfacing. Most sites never do that exercise, and they should.

A real file you can copy

Here's the actual llms.txt for this blog (slightly abbreviated for length):

# AI Marketing Focus
> Practical AI + digital marketing tactics from a 15-year veteran.
> New articles in English and Chinese every week.

## Latest articles
- [LLMs.txt for Digital Marketing](https://aimarketingfocus.com/posts/llms-txt-digital-marketing/): A 2026 practitioner's take on the LLMs.txt spec
- [From SEO to GEO](https://aimarketingfocus.com/posts/geo-aeo-from-seo-perspective/): A practical guide to generative engine optimization
- [AI Search Is Eating Your SEO Traffic](https://aimarketingfocus.com/posts/ai-search-traffic-shift/): What actually works when AI overviews take your clicks

## Core categories
- [SEO](https://aimarketingfocus.com/category/seo/): Search, AEO, content audits
- [Paid Media](https://aimarketingfocus.com/category/paid-media/): Meta, Google, PMax, RSA
- [AI Tools](https://aimarketingfocus.com/category/ai-tools/): Workflows with Claude, ChatGPT, Gemini
- [Marketing](https://aimarketingfocus.com/category/marketing/): Funnels, email, growth, retention

## Optional
- [Archive](https://aimarketingfocus.com/archive/): All 170+ articles, indexed by date

Two design choices worth calling out: the ## Optional section contains the archive because an LLM under context pressure doesn't need every post — only the curated top-of-funnel paths. And every link has a description written in plain prose, not stuffed with keywords. That's the actual format.

Where it fits in the stack

LLMs.txt sits one layer above content and one layer below crawl-control. In a 2026 marketing stack it ranks behind, in this order:

1. Actually good content — the AI cites what's useful, not what's discoverable.
2. Structured data (schema.org/JSON-LD) — feeds entity-level signals that retrieval systems actually consume.
3. GEO/AEO practices — clear claims, citable stats, authoritative author markup, real first-hand experience. (See my GEO guide.)
4. robots.txt AI bot controls — block or allow GPTBot, ClaudeBot, etc. per your licensing posture.
5. llms.txt — a polite description of your best content for any system that chooses to read it.

Notice where it lands: useful, cheap, defensive, but not load-bearing. If you skip everything else on this list and ship an llms.txt, you've written a nice Markdown file. If you do everything else and skip llms.txt, your AI search performance is basically identical.

What I'd actually recommend

Tiered, based on site type:

• Enterprise publisher or documentation site: Ship llms.txt today. Use llms-full.txt only for content you want widely indexed. Review quarterly.
• E-commerce or SaaS marketing site: Ship a thin llms.txt with your top 20 product/category pages. Skip llms-full.txt.
• Small business or personal blog: Skip it for now. Focus on the four layers above. Come back when (and if) an engine confirms it consumes the file.
• Anyone publishing under a paywall or licensing-sensitive content: Do not publish llms-full.txt. The convenience isn't worth the corpus leak.

The single biggest mistake I see right now is teams treating LLMs.txt like a ranking factor and deprioritizing the actual content and structured-data work that drives AI citations. The file is a finishing touch, not a foundation.

If you build a llms.txt for your site, deploy it, and watch for 60 days, you'll probably see what I saw: nothing measurable. That's not a reason to delete the file. It's a reason to keep your expectations honest and your GEO budget focused on the layers that actually move the needle.