Roxana Stingu

So many posts about making your content "machine readable" for LLMs. CEOs of the prompt tracking platforms going on podcasts saying "for the first time we're not writing for humans, we're writing for machines" With the caveat that I recommend, as I always have, that content is well structured and clear (because that's just good writing), it's all backwards if you ask me. The whole point of LLMs is that they're very good at modelling natural language. They create structure from (and find meaning/patterns in) unstructured text. It's literally what they do. That was the breakthrough. That's why we're here talking about it. That's how you can talk to your computer. That's the "magic". Where there's a structure problem (if there is one) is literally at the retrieval/search/parsing layer. Since inference is expensive and context windows are finite, the LLM isn't going to see your full document in a wide retrieval (i.e. a web search). The bit you're optimizing for is making sure that what's retrieved from your document is relevant to the query (which is maths and heuristics before the LLM sees anything). But that's got nothing to do with the LLM not being able to "understand" and "read" your text. It's optimizing for retrieval. For search. The thing we've been doing for decades. If anything, there’s an argument that it’s easier now because the LLM can work with meaning, not just strings. And yes, sometimes there are real structure problems. Tables can be hard to extract. JavaScript-rendered content can be missed. PDFs can parse badly. Crawlers can be blocked. Markup can be malformed. Chunks can be selected poorly. But those aren't "the LLM can’t understand your writing" problems. They're crawling, parsing, extraction, indexing, and retrieval problems. Of course you want to make sure your text doesn't contradict itself, that facts and claims are consistent, and that you get your points across. Externally you want to make sure your brand messaging is consistent. But that's got nothing to do with optimizing for machines. You should have been doing that for humans anyway. Perhaps you get punished a little more for it now, but it was always a problem that you should have been working to solve. If someone tells you that you need to add schema for an LLM to understand your article, you should seriously question their understanding of how LLMs work. If someone tells you that to do "GEO" you need to split your content into chunks with headings, then go and tell them to read any SEO guide published over the past 30 years. We have always been optimizing for both machines and humans. We still are. But perversely, despite the ocean of LinkedIn infographics, the LLM bit is actually more closely aligned to the human bit. The machine bit is the retrieval. That's what you're optimizing. That's what we've always been optimizing. And that's what we'll continue to optimize.