GEO, Getting Cited by AI Engines

SEO plays out on a results page: a list of ten links where your job is to reach the top three. An AI engine hands over no list. It writes one answer, names two or three sources, and stops. No fourth position collects leftover clicks, so an uncited page is simply absent.

The second difference is more technical. An AI engine slices a page into small blocks, stores them, then recalls only the piece that matches the question. Your page competes not as a page but as a pile of paragraphs, each of which must still make sense once pulled out of place.

Before anyone discusses content structure, we confirm the page can be fetched. Many sites are single page apps whose content is assembled in the visitor's browser. A human browser runs the JavaScript, so the screen looks full. Most AI crawlers do not, so what reaches them is an empty shell holding one div and one script file.

The second cause sits outside the code: bot rules at the CDN or firewall. Plenty of hosting panels ship a managed robots.txt that refuses AI crawler user agents, and the owner never finds out because the site behaves normally for humans. The check is short, and you can run it yourself with the steps below.

Because the engine recalls fragments, each fragment is written as an answer that stands alone. Headings follow questions people genuinely ask, and the first two sentences answer them outright with a number or a firm statement. A warm up paragraph that circles the point is the fastest way to get a block skipped.

The commonest writing fault is the backward reference, since a line like as explained above leaves a block incomprehensible the moment it is lifted. Beyond sentences, one table of real numbers gets quoted more often than three paragraphs explaining the same figures, and JSON-LD helps an engine confirm your name, address, and prices do not contradict each other. Our address in Gunung Putri, Bogor is written identically on every page.

Sites clients bring us rarely fail from a shortage of content. Usually there is plenty, arranged so that no single fragment is worth recalling. The failure patterns repeat, and nearly all are fixable without touching the design.

The costliest is the giant page answering twelve questions at once. It reads comfortably top to bottom for a person, but each fragment is half an answer, so none is complete enough to quote. Split it into one question per page and each fragment stands as a whole answer, which finally gives the engine a piece worth recalling. The logic is plain: an engine pulls a part, never the whole, so the part is what has to be complete.

The measurement rests on three sources you can inspect yourself. First the server log, which shows which AI crawlers arrive and which pages they take. If none appears in a month, content work is premature and the problem is access. Second, referral traffic from AI assistants, which lands in analytics as its own source.

Third, a prompt panel: forty questions your actual buyers ask, re run monthly with identical wording, logged against whether you were named and which page was cited. The wording has to stay identical month to month, because once the questions drift you can no longer tell whether the answer moved or your prompt did. A catalogue broken out per product follows the same logic: every product gets a page that can be cited, rather than one list page mentioning all of them in passing.