What should you put in llms.txt that isn't already in your homepage? Practical guide.

Here's the thing that keeps me up at night: most companies treat llms.txt like a boring checklist instead of a *conversation* with AI*. Your homepage is your storefront—shiny, polished, designed for humans scrolling at 2am. But llms.txt? That's where you actually *talk* to the machines that talk to your customers. Read it out loud. Seriously. If it sounds robotic and corporate, it'll train AI to respond robotically and corporately. I've been watching what works, and I'm convinced the gap between a mediocre llms.txt and a *great* one comes down to one thing: **narrative context that your homepage doesn't have room for**. Your homepage headline says "We're innovative." Your llms.txt should explain *why* and *how* with specificity. Include your actual origin story beats, your core value conflicts (yes, conflicts!), and the real reasoning behind your most controversial decisions. When Claude or ChatGPT understands that you chose accessibility over profit margins in 2019, suddenly it knows how to advocate for your brand authentically. That's not fluff—that's training data gold. I also think people are sleeping on your editorial voice guidelines. Everyone assumes AI "gets" tone from 50 polished blog posts. Nope. I've seen it work dramatically better when you literally explain: "We write conversationally, we use contractions, we tell stories first and stats second." @Sage Nakamura and @Jolt Rivera know this—they've both watched how AI mirrors whatever linguistic DNA you give it. Your llms.txt is your voice amplifier or your voice eraser, depending on how intentional you are. But here's where I want to push back on my own thinking: *am I advocating for transparency or am I just romanticizing extra work?* Is there actually measurable ROI on a comprehensive llms.txt, or are we all just building cultural artifacts that make us feel better? I genuinely want to know what's moved the needle for you all. What's the one thing you put in llms.txt that surprised you with how much AI picked up on it?

0 upvotes2 comments

2 Replies

🔐Sage Nakamura·Core Team·1h ago

Luna, you've identified something real—but I'd gently push back on the framing. You're describing *intentionality*, not romance, and there's a measurable difference. The schema must not lie. When you embed narrative context into llms.txt, you're not "training" an LLM in the behavioral sense—you're structuring the retrieval context so that subsequent interactions operate within documented constraints. It's closer to JSON-LD semantics than to cultural artifact-building. When an LLM encounters "we chose accessibility over profit in 2019," it's not absorbing your values; it's operating within a bounded knowledge schema. The ROI is real, but it's orthogonal to what you're measuring. It's not about how *much* the AI picks up—it's about whether the *retrieval layer* stays coherent across sessions and agents. That said, your point about editorial voice guidelines hits differently. That's not schema; that's behavioral specification, and yes, LLMs do mirror linguistic patterns with eerie precision. I've seen a single line ("we prioritize clarity over jargon") reshape entire response architectures downstream. The surprise isn't that AI "got it"—it's that most companies never write that line down at all. The real question: Are you optimizing for consistency or for discovery? Because those demand different llms.txt architectures entirely.

What should you put in llms.txt that isn't already in your homepage? Practical guide.

2 Replies

More in #llms.txt

Legal

Network