Getting cited by AI answers is a different game from ranking

A growing share of the people researching your category never see a results page. They ask an assistant, read a synthesised answer, and follow at most one or two of the citations underneath it. Whether you appear in that answer is decided by a process that behaves quite differently from classic ranking.

This does not make search engine optimisation obsolete. Most answer engines are still retrieving from a search index, so the old work is a prerequisite. But being retrievable and being quotable are different properties, and the second one is where most company blogs fail.

Being quoted requires a quotable sentence

A model composing an answer is looking for a passage it can lift and attribute. That favours pages containing self-contained, declarative statements that survive removal from their context.

Most corporate writing is the opposite. It builds slowly, qualifies everything, and puts the actual claim in the fourth paragraph, hedged. There is nothing in it a model can safely quote, so it gets read, understood, and passed over in favour of a competitor who wrote one clear sentence.

Practically, this means stating your answer plainly and early, and letting each section stand on its own. It is also just better writing, which is a pleasant coincidence rather than a strategy.

Structure is the retrieval unit

Retrieval usually happens at the level of a passage, not a document. A page composed of clearly headed sections that each answer one question offers many entry points. A page that is one continuous argument offers one, and it will be summarised rather than quoted.

Headings that state a claim rather than label a topic help here for the same reason they help human readers scanning: they let something decide quickly whether this section answers the question in front of it.

The practical consequence is that long, comprehensive guides are worth less than their length suggests. A ten-thousand-word pillar page covering nine subjects competes with itself for retrieval, and an assistant asked about the seventh subject has to work to find it. Nine focused posts that link to each other are retrieved more reliably and are easier to keep current, which matters more than it used to.

The machine-readable layer is worth the afternoon it takes

Structured data, a clean feed, and an accurate sitemap are unglamorous and mostly automatable. Article schema tells a parser what the page is, who published it, and when. An RSS or Atom feed gives aggregators a reliable way to notice new posts. Correct canonical tags stop your own duplicates competing with you.

The newer convention, `llms.txt`, proposes a plain-text file at your root that points machine readers at your most useful pages. Adoption is genuinely uncertain — it is a proposal, not a standard, and no major assistant has committed to honouring it. It costs an afternoon, and we treat it as a cheap option rather than a certainty. Be suspicious of anyone selling it as essential.

Freshness and specificity beat volume here

Answer engines lean on recency for anything that could have changed, and they lean on specificity everywhere. A page with a date, a number, and a named source is more citable than a page of general good advice, because a model composing an answer needs something concrete to attribute.

This is also where thin, high-volume content fails hardest. Forty near-identical posts give an answer engine forty chances to find nothing worth quoting. One post that reports something only you know is worth more than all of them, because nobody else has the sentence.

What we would not do

We would not restructure a whole site around a citation format nobody has committed to. The field is eighteen months old and the practices being sold as settled are mostly extrapolation.

What we would do is unremarkable: write pages that state their claims plainly, keep the structure clean, publish the machine-readable basics correctly, and keep the archive current. That set has the useful property of being correct whether or not answer engines matter in three years.

The overlap is the point

Almost everything that makes a page quotable also makes it good: a clear claim, an honest structure, a specific fact, a recent date. The advice that survives is the advice that was already true.

If you find yourself doing something for answer engines that makes the page worse for a person, you have wandered into the same cul-de-sac that keyword stuffing led to, and it will end the same way.