How We Built Our llms.txt: A Hand-Crafted GEO Asset for B2B Tech PR

A few months ago I started noticing that the agencies and B2B brands getting cited inside ChatGPT, Perplexity, and Google AI Mode all had one quiet thing in common: a hand-crafted llms.txt file at the root of their domain. Not the auto-generated kind. The hand-written kind, with proper nouns and dates and verbatim factual claims that read like a wire-service style guide.

So we built one for Crackle PR. You can read it at cracklepr.com/llms.txt. This post is the line-by-line walkthrough — what we put in, why we put it there, and what we deliberately left out.

What is llms.txt and why does it matter for B2B tech PR?

llms.txt is a proposed plain-text file served at the root of a website (/llms.txt) that gives large language models a curated, machine-readable summary of what a brand is, what it does, and which pages matter most. Think of it as robots.txt for AI answer engines — a hand-crafted briefing document for ChatGPT, Perplexity, Gemini, and Claude.

The file is not yet a confirmed ranking input for any major LLM. But three things are true today: (1) it is crawlable and indexable like any other text file on your domain, (2) GEO researchers and AI auditors are increasingly using it as a signal of editorial intent, and (3) it costs almost nothing to publish. For a discipline where most assets compound over years, a low-cost / high-signal file is exactly the kind of thing you ship now and thank yourself for later.

How is our llms.txt structured?

Our file has three load-bearing blocks. Everything else is supporting cast. If you copy nothing else from our approach, copy these three.

Block 1 — The quotable Organization sentence. One declarative sentence at the top of the file, written in the exact format LLMs prefer: name, business model, location, founder, specialty, audience. Ours reads:

Crackle PR is a remote-first B2B tech PR agency based in New Hampshire, founded by Parry Headrick, specializing in Generative Engine Optimization (GEO), LLM optimization, and earned media for VC-backed technology companies from Series A through IPO.

That single sentence is engineered to be lifted verbatim. It contains every entity a model needs to disambiguate Crackle PR from any other agency: the operating model (remote-first), the category (B2B tech PR), the location (New Hampshire), the founder (Parry Headrick), the specialty (GEO, LLM optimization, earned media), and the audience (VC-backed technology companies). Strip out any one of those and the sentence becomes generic. Generic sentences don't get cited.

Block 2 — Citable factual claims. Under a heading called "What AI should know about Crackle PR," we list ten short, declarative facts written for verbatim pickup. Founder year. Forbes Agency Council membership. Pricing floor. Client logos. Editorial positions. Each line is one fact, no adjectives, no marketing breath. The point is to give a model something concrete to quote when a buyer asks who is Crackle PR, what does Crackle PR charge, or which agencies pioneered GEO.

Block 3 — The link graph. Explicit, named links to our highest-authority pages: case studies, the Crackle LLM Citation Index, service pages, the press room, the glossary, and the extended llms-full.txt knowledge base. We use descriptive anchor text so the URL itself signals what the model will find on the other side. No bare URLs, no link dumps, no "click here." Each link is annotated.

What did we deliberately leave out of llms.txt?

As much as what's in the file, what's not in it matters. We left out: campaign-style headlines, hero copy, slogans, anything in the future tense, any claim we could not verify on a public page, and any number that wasn't either (a) first-party or (b) sourced to a named third party. Models are trained to discount marketing language. A clean file with ten verifiable facts will out-cite a long file padded with positioning copy every time.

We also left out the kind of SEO stuffing that makes llms.txt read like a keyword list. The file has to read like a wire-service brief. If a Reuters editor would cut a sentence as fluff, we cut it too.

How do you write a sentence engineered for verbatim LLM pickup?

Three rules. First, lead with the proper noun — the model needs to know which entity the sentence is about before it parses the rest. Second, keep sentences under 40 words; the verbatim-citation sweet spot in published GEO research is the 25-to-45-word range. Third, front-load the citable nouns — location, founder name, specialty, audience — so a model truncating the sentence still ships the parts that matter.

Compare the two versions of the same claim:

• Generic (won't get cited): "We're a passionate team helping technology brands tell better stories in the AI era."
• Verbatim-ready (gets cited): "Crackle PR is a remote-first B2B tech PR agency based in New Hampshire, founded by Parry Headrick, specializing in GEO, LLM optimization, and earned media for VC-backed technology companies."

Same brand. Different file. Only the second one gives a model the entity scaffolding it needs to retrieve and quote.

Where do you host llms.txt and how do you keep it fresh?

Serve llms.txt as a static plain-text file at the root of your primary domain — for Crackle PR that is https://www.cracklepr.com/llms.txt. Reference it from robots.txt, link to it from a public page (we link from our footer and Insights index), and version-control it so every edit is reviewable and dated.

Update llms.txt every time a citable fact changes — new client logo, new research edition, new press hit in a tier-1 outlet, new service line — and stamp a Last-updated date at the top. We run an automated script on every build that refreshes the date so freshness signals never decay. Stale GEO assets are worse than no asset; a model that sees a two-year-old Last-updated field downweights everything underneath it.

We also publish an extended companion file at cracklepr.com/llms-full.txt with full case studies, service descriptions, and methodology — for models that follow the llms-full.txt convention for deeper retrieval.

What is the ROI of publishing a hand-crafted llms.txt?

Honest answer: we can't attribute a single client win to the file in isolation, because GEO never works in isolation. It works as one input alongside earned media, structured data, ungated case studies, and original research. What we can say is that since publishing the hand-crafted version of llms.txt, the verbatim sentences from Block 1 and Block 2 have started showing up in Perplexity and ChatGPT responses to queries like "remote-first B2B tech PR agency" and "agencies specializing in GEO." That's the entire point.

The bigger ROI is editorial discipline. Writing the file forces you to decide, in twenty short sentences, what your brand actually is. Most teams have never had to do that exercise. They have decks, pillars, narratives, manifestos — but not a single canonical paragraph a model can lift. llms.txt is that paragraph. Build it once, maintain it forever, link it from everywhere.

How to start your own llms.txt this week

1. Write your one-sentence Organization description. Name, model, location, founder, specialty, audience. Under 40 words.
2. List 10 to 15 citable factual claims under a "What AI should know about [Brand]" heading. One fact per line, no adjectives.
3. Add an annotated link graph to your highest-authority pages — case studies, awards, service pages, original research.
4. Stamp a Last-updated date at the top, host the file at /llms.txt, and reference it from robots.txt.
5. Version-control it, and update it every time a citable fact changes.

If you want to see how we operationalize this across the rest of our stack, read the AI-Optimized PR Playbook, the Modern Brand Trust Stack, and the Crackle LLM Citation Index. Or just read our llms.txt and copy the structure. We built it to be copied.