First, how we measure this
Everything in this article comes from live checks, not intuition. We send real buyer questions to the engines through their APIs, capture the verbatim answers and their citations, and count a brand only when it is named in the answer text or its domain appears in the sources. Nothing inferred, nothing estimated.
That standard matters because it removes the comfortable explanations. When a brand is absent from 74 of 75 answers, it is not bad luck. Something structural is causing it, and the structural causes turn out to be surprisingly consistent.
Reason one: the engine cannot read your site
Before an engine can recommend you, it has to be able to read you. This fails more often than anyone expects, and at companies that would never believe it applies to them.
The most common version: a site built as a client-side app that serves the same HTML shell for every route. We audited one marketing platform this summer whose pricing, features and case studies pages all returned byte-identical markup to a crawler. A person with a browser sees four different pages. GPTBot sees one page, four times. To every AI engine, that company's pricing does not exist.
The fix is unglamorous: server-side rendering for anything a buyer might ask about, real titles per page, and structured data that states plainly what the company does, where, and for whom. It is a weekend of engineering that outperforms months of content.
Reason two: nothing on your site answers a question
Engines assemble answers from pages that contain answers. Most company websites contain positioning instead: leading, integrated, world-class, trusted by discerning clients. Ask yourself what question that text answers. The engine asks the same thing, finds nothing liftable, and moves on to a site that states its prices, its formats, its timelines and its customers in plain declarative sentences.
The test we apply in audits is simple: take the ten questions buyers actually ask in your category and look for the page on your site that answers each one directly. At most companies we audit, the honest count is zero. The winning competitors usually have a page per question.
Reason three: you are absent from the places engines actually cite
In service categories especially, engines do not primarily cite brand homepages. They cite the layer above: directories, review platforms and listicles. In one 75-answer sweep we ran across an events category, the most cited domains were one incumbent agency and a run of directories and review sites. The brands being recommended were, to a large degree, the brands those third-party pages listed.
This is uncomfortable because it means part of your AI visibility is not on your own website at all. If the five most cited sources in your category do not list you, the engines are choosing from a menu you are not on. Getting listed, properly and honestly, is often the fastest single move available.
Reason four: content volume without entity clarity
The most instructive failure we measured this year: a company in our own category with roughly sixty published blog posts, covering every keyword in the space, that appeared in zero of the 38 commercial answers we tested. Sixty posts, no citations on anything a buyer would ask.
Volume is not the variable. Engines need to understand what you are as an entity: a name that appears consistently, on your site and off it, attached to a clear category, real customers and checkable facts. A company that publishes daily but is described nowhere else on the web is, to an engine, a rumour. A company described consistently in ten places it trusts is a fact.
And one finding that surprised us: being cited is not the same as being recommended. A real estate firm we measured had its own pages pulled in as sources on 31 of 57 questions, and was recommended on 8. The engines were reading its content and using it to recommend its competitors. Content quality had done its job; entity authority had not.
What to do, in order
Measure first. Run the actual questions your buyers ask through the actual engines and record who is named. Everything else follows from knowing which of the four failures above is yours, and most companies guess wrong.
Then fix legibility before content: rendering, schema, per-page titles. Then build one page per buying question, written to be quoted. Then get onto the third-party surfaces your category's answers actually cite. Then re-run the same questions on a schedule, because the only meaningful metric is the share of answers that name you, and it moves slowly and then all at once.
None of this is secret. It is just work, done in the right order, against measured facts instead of guesses.