AI optimization is the new SEO — and large language models (LLMs) are rapidly becoming the interface between people and software.

When someone asks ChatGPT how your product works, how to configure it, or whether it supports a specific feature, the answer is only as good as the documentation it was trained or grounded on.

As a Content Lead here at GitBook, I’ve been focused on improving our docs over the past few months — with a focus on improving AI readability.

In this post, I’ll explain how we structure and write our docs so LLMs can accurately understand how GitBook works — without compromising on the human experience. I’ll also share some practical advice you can apply to your own docs, based on successes we’ve had.

Like all documentation, this is a work in progress — we’re constantly refining it as AI systems evolve. But here’s a snapshot of our current best practices when it comes to AI optimization.

Why GEO matters

Generative engine optimization (GEO) is about making sure AI systems can accurately discover, interpret, and reference your product.

If your documentation is vague, inconsistent or incomplete, AI tools will fill in the gaps. They might guess, or generalize from competitors — but either way, you lose control of the narrative.

Good documentation reduces hallucinations and improves discoverability. In 2025, AI-driven readership of documentation published using GitBook increased by 500% and now accounts for 41% of all docs readers. In other words: AI is no longer a niche audience — it’s a primary consumer of your docs content.

If you want AI systems to recommend your product confidently and correctly, your docs need to be well-written and carefully structured.

How we optimize our docs for AI at GitBook

Most of what we do isn’t AI magic — it’s just disciplined documentation practice. Here are a few of the practices that have brought us success:

Clearly define terms

One of the first pages on our docs site is Concepts. This page covers the most important, product-specific language — things like “space,” “change request,” and “site section.” LLMs struggle when terminology is ambiguous or inconsistently used, so a canonical glossary anchors their understanding and reduces confusion.

Use a strict hierarchy

Every page follows logical H1 → H2 → H3 formatting. We use descriptive headings, concise paragraphs, and structured lists where appropriate, and any code blocks are labeled and contextualized. This predictable structure makes it easier for models to parse meaning and relationships between concepts.

Make pages self contained

While we always aim to avoid duplication, we also try to ensure that each page explains its core concept clearly without relying on excessive context from elsewhere. LLMs don’t always ingest an entire site at once — they often retrieve and reason over individual chunks.

Of course, the other challenge here is keeping things self-contained without every page becoming too long. An LLM’s context window can be limited to a certain character count, meaning long pages can get truncated. So be careful about overstuffing pages — it may hinder rather than help.

Use internal links regularly

While keeping pages contained is important, it’s also vital to connect related concepts so that both humans and AI systems can follow the product’s conceptual graph. A well-linked documentation set teaches the model how features relate to each other.

The good news (for me) is that relative links in GitBook auto-update even when we move pages around, and broken links from external sources are automatically flagged in the UI.

Maintain a changelog

Our changelog features all the key releases and latest updates about the product. And its freshness and clarity are key. Outdated docs don’t just confuse users — they propagate incorrect answers into AI systems.

Always add metadata

Clear titles and meta descriptions are table stakes, but not just for search engines. They provide high-signal summaries that help models quickly understand what a page covers.

The GitBook platform is built for AI ingestion

While these writing practices all help our content perform well, the platform used to publish your docs also matters.

GitBook includes built-in AI-friendly features such as llms.txt and llms-full.txt, which provide structured entry points for models to understand your documentation corpus. Docs published on GitBook also automatically get an MCP server to make it easier for AI systems to interpret your product knowledge programmatically.

Plus, every page can be exported as clean Markdown, making it simple for teams to manually provide high-quality context to an LLM when needed.

As AI is becoming a primary interface to your product, you need infrastructure designed for that reality. We’re hearing from more and more teams using custom-built docs platforms that are rushing to add AI features, and struggling to properly integrate them.​

We believe documentation shouldn’t require retrofitting to be machine-readable — it should be built that way from the start.

This is a work in progress​

One really important thing to keep in mind when optimizing documentation for AI? It doesn’t mean sacrificing clarity for humans.

More than half of your docs readers are still human. So the goal isn’t to write robotic content for AI — it’s to write precise, well-structured, unambiguous documentation that serves both audiences.​

And we’re in an ongoing process to find that balance internally. We’re continuously refining our structure, terminology, and linking strategy as we learn more about how AI systems consume content.

If you’re serious about helping AI understand — and accurately represent — your product, start with your documentation. And if you want a platform designed to make that easier, try GitBook for free and see the benefits for yourself.

→ Get started with GitBook for free​

→ skill.md explained: How to structure your product for AI agents

→ Read the docs