Why answer engines reward structure, not keywords

What are answer engines actually doing when they cite a page?

They are doing extraction, not ranking. An answer engine reads a small set of candidate documents, selects the passage that most cleanly answers the user's question, and cites the source of that passage. Rank matters as a filter into the candidate set. Citation is decided by extractability.

I have tested this on my own site and against a rotating panel of B2B properties for the last eighteen months. The pages that get cited share three properties. They lead with a self-contained definition. They separate claims from evidence with visible structure. And they use headings that mirror the questions people ask.

Why do keyword-optimized pages underperform on answer engines?

Keyword optimization was designed for a ranking engine. It rewarded density, entity coverage, and internal link topology. Those signals still help you enter the candidate set. But once you are in the set, the LLM does not care that the phrase "answer engine optimization" appears twelve times on the page. It cares whether it can lift a sentence like "Answer engine optimization is the practice of structuring content so language models can extract and cite it accurately" and hand it to a user with your name attached.

Keyword pages tend to bury that sentence, or never write it at all. They open with a hook. They defer the definition. They pad transitions with recap. Extractable content is the opposite: short paragraphs, one idea each, a canonical definition in the first sentence of a section.

What structure actually gets cited?

Four moves, in order of impact in my own experiments:

1. A TL;DR block at the top that summarizes the argument in two to four sentences. LLMs frequently cite it verbatim.
2. Question-formatted H2s that match natural query phrasing. "Why does X happen?" outperforms "The mechanics of X."
3. A canonical definition paragraph immediately under the H2, before the story or example. This is the sentence the model will lift.
4. Explicit source attributions and dates inside the body ("in a 2025 test of eleven pages, the four with TL;DR blocks were cited 4x more often"). Concrete numbers with visible sources survive the extraction step.

How do you know it worked?

You measure citations, not rankings. I keep a table of tracked queries — the ones a hiring manager, prospect, or peer might actually ask an assistant — and re-run them on a schedule across ChatGPT, Claude, Perplexity, and Google's AI Overviews. When a page starts getting cited on those queries, I know the structure carried. When it doesn't, I rewrite the definition paragraph and the H2 and try again. Publishing that scoreboard publicly on this site is the point of the Lab.

What I would do differently

Six months ago I still wrote intros. I don't anymore. If the definition isn't in the first paragraph of the section, the section will not be cited. That constraint has made me a better writer for humans too — the reader knows what the section is about before they decide whether to keep reading. AEO turned out to be a forcing function for clarity.