How to measure whether AI engines mention you

A keyword rank measures one thing: where a URL lands in a ranked list for one query. Answer-engine visibility is a different event. A model writes a synthesized answer to a full question. It either draws on and cites a page, summarizes the page without a visible citation, or ignores it in favor of other sources. There is no ranked list to check a position in. There is a generated answer, and the question is whether a specific source shows up in it.

So the tracking method has to change too. Instead of one keyword and one rank, the unit is a full question, phrased the way a buyer would ask it, run against a model. The result is logged in four parts: was the page cited, was it summarized without a citation, was it absent, and what did the answer around it say.

Start with real questions a buyer in the target audience would ask an AI assistant. Not keyword fragments. "What CRM lets an AI agent write to it directly" is a question someone might type into Claude or ChatGPT. "AI-native CRM" is a keyword phrase. It is useful for traditional SEO, but nobody types it into a conversation. The question set should mirror the questions the buyer has at each stage of considering the category, not only the head-term version of the topic.

Run that same set across more than one engine. Different engines pull from different sources and format citations differently. Some show an explicit list of sources. Others weave a source into a plain-text summary with no citation marker at all. For those, "was this page used" is harder to measure. The method has to be engine-specific in places: check for a visible citation where one exists, and compare the summary's claims against a candidate source where it does not.

A single run is a snapshot of what one model version said on one day. Models get updated, and outputs vary even between runs of the same version. A real trend needs the same question set logged repeatedly over weeks or months, the way a rank tracker logs keyword positions. Keep the question wording constant between runs so the comparison is fair.

Across the engines that show citations, some properties of cited pages recur often enough to be a working theory. Not a guaranteed formula. A page that answers one specific question directly, in its first few sentences, is easier for a model to lift a usable summary from than a page that builds up to the answer after several paragraphs. A table with clear headers and units is easier to cite accurately than the same facts in prose. A named, dated source for a claim gives a model something concrete to attribute, rather than an assertion it would have to summarize without backing.

That reasoning is why this page, and the rest of this site, opens with a direct-answer block instead of narrative buildup. A reader, or a model summarizing for a reader, gets a complete quotable answer without reading the rest of the argument first. Whether that structure correlates with being cited more is exactly what the repeated, logged question-set method is for. A theory about what gets cited is only as good as the tracking that can confirm or disprove it over time.