How AI Search Engines Choose Which Sources to Cite, and How to Be One
Something changed in how business buyers research over the last two years, and most company websites have not caught up with it. A meaningful share of the questions that used to go into a search box now go into an AI assistant: ChatGPT, Google's AI overviews, Perplexity, Copilot, and whatever a given buyer's company has rolled out internally.
Those tools do not return ten blue links. They return one written answer, and they cite a small number of sources for it. If your page is one of them, you are in the conversation. If it is not, you are invisible to that buyer, no matter how well you rank in the classic results below.
This article is about what we have learned, from watching which of our clients' pages get cited and which do not, about how to be on that short list.
What the assistants are actually doing
The mechanics vary between tools, but the shape is consistent. When a user asks a question, the assistant runs one or more searches behind the scenes, reads the top results it can access, pulls out the passages that answer the question, and writes a response with citations pointing back to the pages those passages came from.
Two consequences follow. First, you still need to rank, because the assistant is reading search results. Everything in our article on ranking still applies. Second, ranking is no longer enough. The assistant is choosing passages, not pages, and it is choosing the ones it can lift cleanly into an answer.
Write in citable passages
The single biggest difference between pages that get cited and pages that do not is whether the answer exists as a self-contained passage.
An assistant looking for "how long does ISO 13485 certification take" wants to find a paragraph that says, in effect, "ISO 13485 certification typically takes six to twelve months for a company with an existing quality system, and twelve to eighteen months without one, depending on audit availability." It can quote that. It cannot quote an answer that is scattered across four paragraphs, buried in a story, or implied rather than stated.
So structure matters more than it used to. Lead each section with a direct answer to the question the heading poses, then elaborate. Put definitions in one sentence. Put numbers in the sentence with the thing they measure. Use lists for steps and comparisons. Write so that any paragraph, lifted out on its own, still makes sense and still answers something.
Be specific enough to be worth citing
Assistants are trained to prefer sources that add information. A page that says "costs vary depending on many factors" gives them nothing to cite. A page that says "in our experience, a starter installation runs between $8,000 and $14,000, with the range driven mostly by whether existing wiring can be reused" gives them a sentence they will use.
This rewards exactly the kind of content that was already good practice: first-hand specifics, real numbers, named conditions. It punishes the hedge-everything style that a lot of corporate content defaults to. If your company knows the answer, say it plainly.
Make the source look trustworthy to a machine
Assistants weigh signals of reliability, and they are similar to the ones people use, expressed in ways a program can read:
- A named author with a stated role, ideally with a short bio and a link to more of their work.
- Dates, both published and updated, so the assistant can tell the information is current.
- Cited sources for claims, linked, so the answer chain can be followed.
- Consistent facts across your site. If one page says a lead time is four weeks and another says six, neither gets trusted.
- Clear identification of who you are, on an about page and in the footer, so the assistant can establish that a real organisation stands behind the content.
None of this is new advice. It is the same advice with higher stakes, because a machine applies it without the benefit of the doubt.
Cover the follow-up questions
When a person asks an assistant a question, they usually ask two or three more in the same conversation. The assistant goes looking for each one. A site that has answered the whole chain, what it is, how it works, what it costs, how long it takes, what can go wrong, tends to be cited repeatedly in the same conversation, which is how a buyer comes to feel that your company is the authority.
This is the same topic-cluster logic that helps with ranking, seen from the other side. Depth of coverage on one subject beats breadth across many.
Make the page readable by the tools
A few technical points, because assistants can only cite what they can read:
- Do not block AI crawlers in your robots file unless you have decided you do not want to be cited. Several of the major tools identify themselves and respect those rules.
- Keep the main content in plain HTML, not inside images, PDFs, or scripts that have to run before the text appears.
- Use structured data where it fits: article, organisation, author, FAQ. It is not required, but it removes ambiguity.
- Keep pages fast and uncluttered. Assistants time out on slow pages the same way people do.
What this means for your content plan
The practical upshot is that the content that gets cited by AI search is the same content that ranks in classic search and the same content that persuades a human reader, written a little more directly. Clear questions, direct answers, real specifics, a named expert, on a site that covers its subject properly.
Companies that already do this are finding themselves cited without changing much. Companies that relied on thin pages and clever keyword placement are finding that the assistants skip them entirely, and that there is no trick to fix it. The only fix is to become the source worth citing.
