AI Citability is the likelihood that an AI system will pull a specific piece of content into a generated answer and cite it as the source. It is a property of the page itself, assessed before any prompt is ever run against it.
For a marketing buyer, this matters because it is the closest thing to a leading indicator in AEO work. Share of voice and appearance rate tell you what already happened across a sample of prompts, which is useful but retrospective and dependent on which queries you tracked. Citability tells you whether a page is built to give it a fair chance the next time a model goes looking for something to quote. If a landing page scores low, no amount of prompt tracking fixes the underlying problem, because the content itself is not shaped for extraction.
In practice, citability is assessed across four dimensions, each measuring a different failure mode. Answer readiness checks whether the page states its answer directly, in the first sentences, rather than building up to it through a long introduction; FAQ-style sections and plain definitions score well here. Structure checks the scaffolding around that answer: heading hierarchy that mirrors how people actually phrase questions, short paragraphs, and lists used where a list is the honest way to present the information. Extractability checks whether a paragraph can stand alone as a complete thought once lifted out of its surrounding context, since models tend to quote self-contained statements rather than sentences that depend on the one before it. Credibility checks the trust signals around the claim itself: named authorship, a visible publication date, links out to sources a model already trusts, and schema markup that makes those facts machine-readable rather than merely human-readable.
Take a product comparison page that scores 42 out of 100. An audit typically finds the answer buried three paragraphs in, no FAQ schema, and a byline with no date. Rewriting the page to open with a direct comparison sentence, adding a marked-up FAQ block and a dated author, is a two-hour edit that can move the score to 78. That is measurable and repeatable. It does not guarantee a model will cite the page tomorrow; retrieval also depends on the query, competing pages available at that moment, and the model's own selection behaviour, none of which the page controls. Citability is a strong predictor, not a promise.
This is also where the confusion with share of voice sets in, because both numbers sound like they measure the same thing. They do not. Citability is computed from the page: crawl it, score the four dimensions, get a number, and that number does not change until the page changes. Share of voice is computed from the outside: run a basket of real prompts against ChatGPT, Gemini, Perplexity and the rest, count how often your brand gets named against competitors, and that number moves with every model update, every competitor's content change and every shift in what people happen to be asking. A page can carry a citability score of 90 and still show up in zero of a hundred tracked prompts, because none of those prompts touched the topic it answers. Conversely, a mediocre page can get cited once by luck of timing. Citability is diagnostic and actionable at the content level; share of voice is the outcome you are ultimately trying to move, aggregated across many pages and many prompts, and it is the number that should show up in a board deck.