Why is content judged by fragment rather than by page?

When an assistant answers a question, it does not read your whole article. A retrieval system splits the content into fragments, finds the ones that resemble the question, and hands the model only those pieces. The model writes the answer from what it received.

The consequence is uncomfortable for anyone who writes well at length: your article competes fragment by fragment, not as a whole work. A brilliant section that only makes sense after reading the previous ones reaches the model without its context and gets dropped.

What makes a section stand on its own?

That it can be read alone and still make sense. Three concrete rules:

  • Repeat the subject instead of using pronouns. “A CRM implementation costs…” survives extraction; “this costs…” does not.
  • A heading in question form. It tells the retrieval system exactly which query that block answers.
  • The answer in the first two sentences of the section, with the development after. If the conclusion sits at the end, the retrieved fragment may not include it.

It is the same inverted-pyramid discipline journalism uses, applied to each section instead of to the article as a whole.

Which kinds of claims get cited most?

Verifiable ones. A model choosing between two sources prefers the one saying something checkable and attributed over the one saying something agreeable.

Compare these two sentences about the same fact:

“Many Colombian companies still have not digitized their operation.” — nothing citable there.

“Only 35% of companies in Colombia run a full ERP; the remaining 65% still work with spreadsheets or isolated systems.” — a concrete figure, a clear contrast, attributable.

The second can be dropped into a generative answer; the first cannot. And the price of the second is that you have to be able to stand behind it: if you cite a figure, cite where it came from too.

What does structured data actually do?

Structured data does not get you cited. It gets you understood without ambiguity, which is a precondition.

What it settles, concretely: that a page is an article and not a product listing, that its author is an identifiable organization, that the questions on that page are frequently asked questions with their answers, and — most important for a brand — that all your pages are talking about the same entity. Without that last point, an engine can treat your trading name and your legal name as two different companies and split what little trust it has between them.

What content gets ignored every time?

Five patterns, in order of frequency:

  • The introduction that does not answer. Three hundred words of context before the first useful claim.
  • Promotional language. “We are leaders,” “innovative solutions.” A model cannot verify it, so it does not use it.
  • Content that depends on JavaScript. Many AI crawlers do not run scripts: if your text is injected in the browser, to them the page is empty.
  • Orphan fragments. Sections that open with “besides,” “on the other hand,” or “this means that.”
  • Content with no date and no author. Two cheap trust signals plenty of people leave out.

Does this mean writing for machines instead of people?

No, and confusing the two is expensive. Text assembled to please algorithms reads badly, and whoever arrives after an assistant cites you is a person who leaves in ten seconds if they do not find what they came for. On top of that, search engines' helpful-content systems explicitly penalize that pattern.

The right formulation is simpler: you write for humans and structure for machines. Everything in this article is structure — where the answer goes, how a section opens, what the markup declares — not padding. A good article well structured wins; a bad article well structured is still a bad article.

How do you check whether your page is extractable?

Two tests that take ten minutes. The first: fetch your page with an AI crawler user agent and check whether the main text is in the raw HTML. If it is not, nothing else matters.

The second: take any section, copy it on its own into a blank document, and read it as if it were all that existed. If it does not hold up alone, no engine will be able to use it. We run both tests on our own site before proposing them to a client.