Do you remember that scene from Wolf of Wall Street where Jordan is hammering home that sales script with those new hires? Showing them word for word how to sell a penny stock?

He was getting people to use a script because when people are new to selling, a script makes them more effective. It gives them the right words and teaches them the correct sequence to close and create a customer.

A sales script is just a guide ideally built by the person who knows how to sell the thing the best, so others can sell the thing just as well as they can.

It’s how to create a 1 to many sales team. As people get better they can bend the rules and go off script, but at first you want your many to match your best one.

Claude, ChatGPT, Gemini etc are now your many new sales team whether you want them to be or not. Your prospects are likely talking to their AI about you before they’re talking to you about you. It’s a conversation happening constantly, for every product, and it’s a channel that can convert—when AI knows what it’s talking about.

So the question is simple: do you trust that they are selling your product as well as you know how to? That they have the right words and sequencing? That they’re not just making stuff up?

If you’re a big company with lots of content online, maybe they do. If you’re new with less of a footprint, almost certainly not. For early stage companies, AI is like a brand new salesman who’s not trained on what they’re selling.

A few weeks ago I wrote about the Ask-AI buttons—those little “ask Claude about us” links you can drop in your footer. A super cheap and easy marketing tool.

What I found for my own app, Little Moments, was that it made stuff up, spoke about journaling in general, and got stuff wrong about me and my unique angle.

When it works, it really works. AI-referred visitors convert somewhere around 4–5x higher than organic search—one study of 312 B2B companies found a 14.2% conversion rate for AI referrals versus 2.8% for Google organic—because they arrive already sold by a source they trust.

The whole idea today follows from that: the AI has become your most persuasive salesperson, and most products are letting it improvise the pitch.

And this is an important one today, because Anthropic is already telling us we need to do this, without telling us anywhere officially.

Everything you’re reading here runs on beehiiv—my newsletter, the website, the analytics, the automations, all of it. I moved my newsletter off Substack because beehiiv treats a newsletter like a real business: it takes 0% of your revenue, hands you a proper API and MCP to build on, and gives you the segmentation and growth tools to run it like a product.

And if you’re a founder, today’s issue is the argument: algorithms—now the AI ones too—decide who gets seen. Your list is the one distribution channel nobody can throttle, rank, or take away from you. Build one next to your product to reach customers, keep telling your story, and win investors; this is where I’d do it.

My code THEDIFF30 gets you 30% off your first 3 months.

🥷 When someone asks an AI about your product, it answers whether you’ve prepped it or not. So write the sales script it reads: a clean llms.txt of the facts, plus honest “us vs them” pages that openly admit where you lose—so its second opinion comes out accurate and fair instead of invented.

There’s one thing you can do in just a few hours (or less), that gets your prospects AI to sell your product the same way you are able to.

Two files, one job. The first is an llms.txt—a plain markdown file at yourdomain.com/llms.txt that tells models what you are, who it’s for, your best features, and where your docs live. The second is a set of honest comparison pages, the kind that say “great for X, worse for Y.”

You can’t control what the AI says, but you can control what it reads and therefore heavily influence what they say.

Llms.txt is the harpoon you give to AI—and it’s not just something early stage companies should be doing…

Anthropic is doing it, which IMO is setting a clear standard that is silly to ignore.

Let’s break it down, but first a quick word from Granola.

My new work sidekick is Granola—the AI notepad with no bot that joins your calls—it transcribes straight from my device audio on Zoom, Meet, Teams, Slack, and even in-person—and quietly turns every conversation into clean notes while I just talk.

I’m big on notes as a PM, but I prefer to listen actively to people without typing/writing on the side. I’ve tried a few of these but Granola is genuinely the best imo.

I’ve been loving their new Chat feature: months of meetings become a searchable memory, so I ask things like “what did I promise the sponsor in June?” and get the answer with citations to the exact note. And its new Briefs feature preps me overnight before external calls—who I’m meeting, what we covered last time, recent news. It’s the closest thing I have to a chief of staff.

When I tell you my app is is great, I’m a salesperson, and you discount it. When an AI the buyer already uses every day explains your product, it reads as a second opinion—and second opinions are the most persuasive. The numbers back it up: Ahrefs found 0.5% of visitors arriving from AI search drove 12.1% of total signups, a 23x multiplier, and Adobe’s Q1 2026 data has AI referral traffic converting 42% better than non-AI traffic—a reversal from 38% worse just twelve months earlier. These people show up having done their comparison shopping inside the chat.

The problem is what the model has to work with. If you’re a big documented brand, it has plenty. If you’re anyone else, it’s reading a JavaScript-heavy marketing site it can barely parse—or it’s guessing. And a guessing model doesn’t stay quiet. It does what Claude did to me: invents a confident paragraph that makes it sound like the founder doesn’t know their own product. That’s worse than no answer.

llms.txt is a markdown map of your site for machines, proposed by Jeremy Howard of Answer.AI in September 2024. Instead of making a model crawl your whole site, you hand it a curated list of what matters.

Why does the AI even know to look there? Same reason every crawler on earth checks robots.txt: it's a convention at a predictable address. Models and AI search tools don't "discover" the file—/llms.txt at your root is a well-known path baked into a growing number of tools (Cursor and other coding agents already fetch it), so when an AI goes looking for ground truth about you, it can grab your whole story in one request instead of spidering through your site hoping to piece it together.

Why markdown? Because models read tokens, not rendered pages. Your actual website is mostly noise to a machine—nav bars, cookie banners, scripts, div soup—and if it's JavaScript-rendered, half of it may be invisible to a crawler entirely. Markdown is pure signal: plain text with just enough structure (headers, lists, links) that models were trained on billions of pages of. Nothing lost in parsing, and far more of you fits in the context window.

And that's why it works like a sales script rather than a pile of content. Your blog posts were written to persuade a human scrolling by—the facts are scattered across pages, wrapped in prose, some of it outdated. A model reconstructing your pitch from that is the new hire winging it. llms.txt is the Wolf of Wall Street script: the right words, in the right order, in one canonical place. You've removed the model's need to guess what matters and what to say first.

A big lever specifically is the honest comparison. A page that says “we’re great for X, and honestly worse for Y than [competitor]” gets cited and believed by AI because it isn’t a brochure. It’s the same thing I found with the Ask-AI buttons: the persuasion came from the answer being balanced, not overly glowing. A model quoting “amazing in every way” convinces no one. A model quoting your fair, specific trade-offs closes the sale. The fairness is the marketing.

There's a mechanical reason for this. Models are trained to sound like a trusted advisor, not an ad—balanced, hedged, "X is better for this, Y is better for that." When a model builds an answer, it reaches for sources that let it be just that. Your honest comparison page is already the exact shape of the answer it wants to give, so it gets quoted nearly verbatim. "We're the best at everything" pattern-matches to marketing copy, and models discount it the same way people do—it either gets ignored or laundered into something neutral you didn't write. Admitting where you lose is what makes you quotable where you win.

I love this idea, but the reality is we don’t know if it works.

There’s some doubt about whether llms.txt does anything today. Trakkr studied it against AI citations and found no correlation—a p-value of 0.85 across 337,000+ citations, statistically indistinguishable from noise.

OpenAI’s crawler honors robots.txt but doesn’t officially use llms.txt, and Anthropic doesn’t officially say its crawlers read the standard.

At this point you might rightfully be thinking…so wtf are you telling me this idea for?

Because I strongly believe—with the case study evidence coming up below as my conviction—that this will very soon be a real standard used.

Plus, no reason to not just ship it. The downside is an afternoon and the upside is real if the models start fetching it. Don’t ship it expecting a spike necessarilly, but treat the file as cheap insurance that’s forward-compatible.

There are no guarantees…but (I think) it can deliver

…big day for Leo references it seems

I say this is worthwhile cheap insurance for you, because a) Anthropic is leading this standard, and b) big companies are noticing and getting ahead

The guys who build Claude serves both llms.txt and llms-full.txt at docs.anthropic.com—a slim index that points to a full markdown export of their docs. AKA the people who understand better than anyone how models read the web decided this was worth shipping.

Takeaway: when the folks building the reader also write the file, it’s table stakes. At the minimum, match them.

Stripe’s world spans payments, billing, fraud, and issuing. Their llms.txt is organized around major product and resource areas—Payments, Checkout, Webhooks, Testing—mirroring their API, so a model can point a developer to the right corner without wading through hundreds of endpoints.

I want to see Stripes example →

Takeaway: structure your brief the way a confused buyer would ask, not the way your org chart looks.

Cloudflare’s llms.txt runs thousands of lines covering Workers, R2, D1, Pages, and 20+ other products. If a company with that much surface area can fit its map in a single file, the three-page site has no excuse.

I want to see Cloudflare’s example →

Takeaway: complexity isn’t the blocker. It’s one file no matter how big you are.

Vercel’s llms.txt is the most thorough one I’ve found out there—it maps their entire documentation structure with a one-line description for every page, so a model always has a labeled door to walk through. More to maintain, but nothing gets left out.

I want to see Vercel’s example →

Takeaway: there’s a spectrum from “curated map” to “full index”—pick the end that matches how much your users lean on AI to use you.

Supabase runs the backend under my apps, and it publishes an llms.txt at supabase.com/llms.txt that acts as an index pointing to separate, language-specific files—a JavaScript reference, a Python one, a CLI one. When your docs are too big for a single file, the index-of-indexes keeps each one focused.

I want to see Supabase’s example →

Takeaway: if one file would be a 500-page dump, make it a table of contents that points to smaller ones.

One markdown file: a one-line description of what you are, who it’s for, your top 3–5 features in plain language, and links to your key pages and docs. Host it at yourdomain.com/llms.txt.

“You vs [the obvious competitor],” and mean it—include where they’re the better pick. This is the part that does the real work today for being cited.

Open an Incognito chat (so there’s no memory of you) and ask “tell me about [your product] vs [competitor].” That’s your before. Fix what’s wrong, then check again. Add “AI assistant (ChatGPT, Claude…)” to your signup survey, and watch for a bump in “direct” traffic—because around 70% of AI referrals get mis-attributed as direct in GA4 anyway.

Once your llms.txt is live, add yours to the llms.txt directory here—free visibility, and it’s where builders (and increasingly, the models themselves) go looking.

A few dos and do nots…

It could be the cheapest afternoon of marketing infrastructure you’ll spend this quarter. Go feed the machine.

There’s also a neat tool that makes this much easier, check it out 👇

› The build-it-yourself guide — Firecrawl’s walkthrough for generating your llms.txt (and llms-full.txt) from a single URL, and what actually belongs in each. → How to create an llms.txt file (8 min)

› How the best files are actually built — a teardown of Anthropic’s, Stripe’s, Cloudflare’s and Vercel’s real llms.txt, with the structural patterns worth copying. → Real llms.txt examples, dissected (10 min)

› The “should you even bother?” take — Ahrefs on what llms.txt does, what it doesn’t, and which crawlers actually read it today. → What is llms.txt, and should you care? (9 min)

› The rabbit hole — a community index of 800+ live llms.txt files, sorted by company, to steal structure from before you write your own. → Browse the llms.txt directory (long lol)

🛠️ A tool I found — Trakkr, the AI-visibility tracker behind the citation study I referenced above. It shows how often ChatGPT, Claude, and Gemini actually mention your brand—and against who—so you can measure today’s play instead of guessing. → Track your AI mentions

📺 A video I loved — “What’s It Like to Be Killed by Nature’s Most Brutal Predator,” by Kurzgesagt. It puts you inside the (very short) final moments of prey meeting nature’s most efficient killer. Gorgeous animation, mild existential dread, and a masterclass in making science unskippable. → Watch it

📬 A read I enjoyed — “Be Plainspoken” by Boz, Meta’s CTO. A two-minute essay on why paragraphs have five sentences when they could have one. It’s basically today’s play applied to humans: plain, honest language is what gets trusted—and quoted. → Read it