There's something that happened to us last year that, if I'm honest, I underestimated at first. We started to see that some articles that used to bring traffic stopped showing up in AI conversations. It's not that they were badly written. It's not that they didn't rank on Google. They simply weren't being cited.
And that changed the question. It was no longer: "How do we rank better?" It was: "How do we get an LLM to choose us as a source?" We went from writing only for humans to structuring content that can also be processed, understood and cited by language models. I'm sharing here the most important takeaways from a practical standpoint.
The real shift: from writing articles to designing knowledge
For years we optimized for the click. Today many answers are resolved without the user ever visiting a website. That means your content has to be able to:
- Answer fast.
- Be clear.
- Be structured so an AI can take it apart and put it back together.
It's not about writing "for robots". It's about organizing what you already know better.
1. Structure: if it's not organized, it doesn't exist
Models need a clear hierarchy:
- An H1 that actually answers a question.
- H2 headings that develop subtopics without illogical jumps.
- Short paragraphs (50–150 words).
- One idea per paragraph.
- A first sentence that works as a summary.
When content is structured this way, AI can extract answers effortlessly. When it isn't, it simply moves on.
2. Clarity: authority without grandiloquence
Something we saw a lot: "interesting" but barely citable texts. Phrases like "Many experts believe…" or "Most companies…" mean nothing to a model. What does work:
- Concrete data. Dates. Explicit sources.
- Clear definitions on first mention.
- Define terms. Avoid unnecessary metaphors. Get to the point.
In generative engines, clarity matters more than creativity.
3. The inverted pyramid applied to AI
Another structure we started using systematically:
- Level 1: The direct answer in the first sentence.
- Level 2 (50–200 words): Why it matters, when it applies.
- Level 3 (200–800 words): How it works, examples, comparisons.
- Level 4: Methodology, expanded data, cases.
LLMs can extract information from any level. But if there's no order, they can't do it accurately.
4. Writing for the machine too (without breaking the experience)
We introduced a "technical layer" to the content: technical summaries at the end, structured data (schema / JSON), high density of entities, dates and numbers. It's not to deceive. It's to make abstraction easier. Always with full transparency and real usefulness.
5. Reverse engineering: understanding what's being cited
A simple exercise I recommend:
- Run 30–50 queries in Perplexity or ChatGPT about your topic (in incognito or a new account).
- Analyze which sources keep coming up.
- See what structure they use: lists of 3 or 5 options, comparison tables, direct definitions, final recommendations.
It's not about copying formats. It's about understanding patterns. Models look for structural consistency.
The mistakes we see most
- Long introductions that say nothing.
- Blocks of text with no subheadings.
- Content with no date.
- Forced CTAs that break the intent.
- Generalizations without data.
AI doesn't penalize. It simply ignores.
Content is no longer just marketing
It's infrastructure. Structure is the API. Data is the currency. Clarity is the speed. It's not about producing more pieces. It's about turning some articles into well-designed knowledge bases.
If your brand wants to exist in the LLM conversation, it has to be easy to cite. And that starts with how you write.
