There is a finding in this category that surprises people more than any other, and most tools cannot report it at all.
Your page is in the assistant's source list. A competitor is in the sentence. The model read you, used what you wrote, and recommended somebody else.
I want to describe why it happens, how to tell it apart from ordinary absence, and what actually fixes it, because the instinctive response is almost always the wrong one.
Two questions that get collapsed into one
Every AI answer contains two separate facts about your brand.
Were you named in the prose a person reads? Were you cited in the sources underneath it?
Those are independent, which gives four outcomes rather than two: named and cited, named but not cited, cited but not named, neither. Most tools in this category report a single mention rate, which merges the middle two into noise and throws away the part that tells you what to do next.
We split them because the work implied by each is completely different, and the four states are reported separately everywhere they appear.
What cited-but-not-named actually means
It means retrieval is working.
That is worth saying clearly, because the emotional reading of "not mentioned" is that the assistant has never heard of you. If your URL is in the citation list, that reading is wrong. The crawler reached you, the retrieval layer selected you as relevant, and the model used your page while writing an answer that recommends a competitor.
You have the model's attention. You are spending it badly.
In practice we see three causes.
Your page describes the category, not you. Educational content ranks well and gets retrieved often. A thorough "what is X and how does it work" page is exactly what a model wants for background. It is also a page with no claim in it. The model takes the explanation and gets its recommendation from somewhere that made one.
Your page describes a competitor more clearly than it describes you. This happens on comparison pages, roundups and glossaries. If you wrote a fair, specific paragraph about a rival and a paragraph of marketing language about yourself, the specific one is the liftable one. You wrote the sentence that recommended them.
Your claim is not in a liftable form. Assistants assemble from passages. A concrete claim split across a pricing table, a feature grid and three footnotes is invisible to that process. The same claim in one plain sentence is not.
Why more content is the wrong reflex
The instinctive fix is to publish more. It is almost always wasted.
The model is already retrieving you. Adding pages gives it more surface from a source it has already selected, which changes nothing about the sentence it writes. The problem is the content of the retrieved passage, not the quantity of retrievable pages.
Rewriting one paragraph on the page that is actually being cited will do more than ten new posts. The hard part is knowing which page that is, which is a measurement problem rather than a writing one. Our citation tracking exists to answer it: which URLs are cited, on which prompts, on which engine, and whether you were named in the same answer.
The pages briefing your category are mostly not yours
This is the uncomfortable structural fact underneath it.
Pew Research Center, examining the sources listed in Google's AI summaries, found that Wikipedia, YouTube and Reddit collectively accounted for 15% of the sources listed, a similar share to their presence in standard results. The rest is spread across the open web, and in most commercial categories a large part of it is third-party comparison content rather than vendor sites.
So even when you are cited, you are usually one voice among several, and the others are describing you. If a roundup post has you as "a solid budget option," that phrase is in the retrieval pool every time someone asks about your category.
Fixing how third parties describe you is slow, unglamorous and one of the highest-leverage things available. It is also why the comparison content we publish names its criterion up front: a vague claim is not liftable, and a specific one is.
Telling it apart from ordinary absence
You cannot see any of this from a mention percentage. You need the answer text and the citation list, stored together, for the same run.
The diagnostic is simple once you have both:
- Cited, not named, repeatedly, on one engine. Framing problem on the cited page. Rewrite the passage that is being retrieved.
- Cited, not named, on every engine. Category-level positioning problem. Your page reads as background material rather than as an option.
- Not cited at all. Retrieval problem. Check robots.txt and indexing before touching the copy. Google's AI features guidance covers the snippet requirement, and its crawler documentation covers Google-Extended; OpenAI documents its crawlers separately by purpose and Perplexity documents PerplexityBot.
- Named, not cited. The model knows you from training and did not read you today. Nothing you publish is holding it up.
Engines differ here, which is why we keep them apart. Perplexity cites more heavily than anything else, so it surfaces this pattern earliest. Claude's web search cites fewer sources per answer, so an absence there carries less information.
An example of the rewrite that works
The abstract version is unconvincing, so here is the shape of it.
A page that gets cited and ignored usually reads like this:
AI visibility tracking involves monitoring how often large language models reference your brand across various platforms. Organisations use a range of approaches, including manual spot checks, third-party tooling and internal dashboards, each with distinct advantages and trade-offs.
Accurate, useful as background, and it contains no claim. A model writing an answer takes that explanation and then needs a recommendation, which it gets from a page that made one.
The version that gets named looks like this:
AIMentionTracker runs all six engines on every plan, stores each answer in full with its citations, and shows the number of checks behind every rate. Below five checks it shows a fraction rather than a percentage.
Same subject, different job. It is specific, attributable, and short enough to lift as a unit. Notice that it is not more promotional than the first version; it is less. It replaces adjectives with facts, which is what makes it liftable.
The test I use: could a model quote one sentence of this page and have said something concrete and checkable about us? If every sentence needs three others around it to mean anything, nothing is liftable.
What to do this week
Take the five prompts you most want to win. Run each one enough times to have a rate rather than an anecdote, on every engine you care about. For each answer, record two things separately: were you named, and were you cited.
Then look only at the answers where the second is yes and the first is no. Open them. Read the paragraph of yours that was cited, and read the sentence that recommended somebody else. The gap between those two pieces of text is the entire problem, and it is usually obvious once they are side by side.
That is the whole method. The measurement rules we hold ourselves to are public, and every plan stores full answer text with citations, because a number you cannot open is a claim you have to take on trust.