Notes from the codebase. And whatever else.

ChatGPT summarises a competitor, Perplexity links someone else, and your own answer to the exact question sits on a page nobody quotes. The instinct is to blame the writing — but often the crawler never saw your text at all. AI crawlers don't run JavaScript (Vercel's data: ClaudeBot downloads JS on a quarter of requests and executes none of it), so a client-rendered page is a blank div to them. Why only Googlebot renders, and why the fix is a CMS decision, not a coding job.

Before you spend a euro on AEO, run one test — it takes about a minute and answers the only question that comes first: when an AI crawler asks for your page, does it get your content or an empty shell? View source, curl the URL, or kill JavaScript, then look for a real sentence of your body copy in the raw HTML. How to read the result, the nuance that hides the problem (Googlebot renders, so Search Console looks fine), and the layer that turns readable into cited.

Add an llms.txt file and the models will prefer you — it's become the default first move, the AEO equivalent of 'add a meta description.' It's also, on the current evidence, mostly a placebo: no major AI company has confirmed it uses the file, and there's no measured citation lift. Worth five minutes, not worth mistaking for the work. The honest status, and the levers that actually have evidence behind them — all of them structural, none of them a text file you paste in.

Growth leads and SEO specialists are running Claude Code now, most without ever opening a terminal. The sourced rundown: workflows marketers have shipped (scheduled lookalike outbound, positioning grading, competitor ad intel, $30/month tools replaced), what skills are and why they make it click, where the raw tool gets hard for a non-developer, and why it all gets convenient when the agent is the core of the CMS rather than a bolt-on.

People wire Claude Code into content sites four ways — an MCP server on a headless CMS, the WordPress REST API, a hand-rolled admin, or plain markdown in the repo. Each has a sharp edge that only shows up in production: MCP tool definitions eating a third of your context, permission prompts between an editor and a typo fix, onboarding that needs a developer, and client-rendered pages answer engines can't read. The sourced rundown — and what changes when the agent is the architecture, not a guest bolted onto one.

The plain-words tour: one of the first LLM-native CMS frameworks, where you edit by talking to an agent. A full CMS ready on day one, a top-tier inline editing flow, near-zero onboarding for content editors, content ops that plug straight in, your own models, and open source.

The problems with running a content site on a headless CMS are old and well-documented. What's new is the alternative: an agent-native architecture where the agent is the primary actor, not a client bolted onto an API built for human editors. Headless still fits the minority it was built for. For everyone pushed onto it past the no-code ceiling, the math changed.