Retrieved but not cited: the most valuable gap in your AI visibility

Most conversations about AI visibility, about whether your brand shows up in the answers of ChatGPT, Perplexity or Gemini, know only two states: your brand appears, or it does not. That is too coarse. The data holds a third state, and it is the most useful of all: the model retrieved your page but did not cite it. Anyone who only looks at “visible, yes or no” misses exactly the places where the least effort moves the most.

Two things that get confused

When you ask ChatGPT, Perplexity or Gemini something, more happens behind the scenes than a block of text. The system first gathers a set of pages it judges relevant to the question. That is retrieval. From that stack it then selects which sources actually feed the written answer and get named beneath it as sources. That is citation. ChatGPT, Perplexity and Gemini handle the details differently, but the basic pattern of gathering and selecting is common to all of them.

The two are not the same, and the difference decides the impact. Being retrieved only means the model considers your page relevant enough to look at. Citation is more. Here your content was convincing enough to become part of the answer the user reads. Content that is retrieved but not cited may shape the answer in the background. But only what gets cited reaches the eyes of the potential customer, as a named and linked brand.

Why the gap is good news

Picture filling a job opening. A page that is retrieved but never cited is the candidate who made the shortlist and then did not get the role. Bad? Sure. What matters more is the other side: she was close. Her relevance is beyond question, or her application would never have landed on the table in the first place.

That is exactly what makes this gap so valuable. If you are not retrieved for a question at all, you often need a new page, new content, sometimes fundamental authority work through third-party sources. The second case is the better one: retrieved but not cited. Then usually only the existing page has to become more convincing. Nothing is missing here in the sense of “something is absent.” What is missing is the citable statement on a page that already exists. The lever is there. It is just lying idle.

Many GEO audits do not separate these two cases. They report low visibility and recommend “more content” across the board. That is expensive and often beside the point. Whoever gets retrieved rarely needs more text. They need the existing text in a form the model can lift out as a reliable statement.

To stay honest: not every gap closes through better wording. Sometimes the cited outside source is simply more authoritative. Then the problem is trust, not word choice, and it resolves through signals from outside rather than through the page itself. The more common case stays the easier one: relevance is established, only the content does not yet convince.

An example from practice

Take an advisory page on an important buying-decision question. It shows up in the measurement data: it is retrieved four times for this question but cited not a single time. Next to it stands a third-party source that is cited consistently for the same question.

Look at both pages and the difference is rarely the length. The third-party source answers the question in clear, liftable sentences, often with a clean FAQ section and an unambiguous, verifiable recommendation. The page of your own circles the topic but stays vague at exactly the point where the model would need a citable statement. The fix is not a rewrite. It is answering the central question directly and phrasing that answer so it can stand on its own.

What makes a page citable

Between retrieved and cited there usually sit only a few concrete properties. The answer to the customer’s question belongs at the top, not at the end of a long text where the model first has to hunt for it. It should stand as a clear, verifiable statement that still makes sense when lifted out of the page. A format that is easy to grasp helps: an unambiguous recommendation, say, or a clean question-and-answer section with precise figures instead of vague paraphrase. Above all: stay concrete. A page that only circles a topic loses to one that names the question and answers it.

How to find the gap

This third state becomes visible only if you measure at the prompt level, that is per individual question, rather than averaging over everything. For each question you need two values side by side: how often your URLs were retrieved and how often they were cited. No ordinary analytics program delivers these; that takes a tool that measures AI visibility specifically, such as Peec.ai. By hand you can only check whether you are cited at all; retrieval itself, and with it the actual gap, becomes visible only with such a tool. A page with high retrieval and low citation is then your priority list, before you write a single line.

I use exactly such a tool; it delivers these values per question, including the third-party sources cited instead. But the method does not hang on any particular tool. Any source works, as long as it gives you three things per question: your visibility, your retrieved and cited URLs, and the strongest third-party sources for exactly this question. Those third-party sources are not decoration. They show which structure and depth the model orients itself by for this question, and with it the bar your page fails to clear.

The values yield a simple sorting. If you are well visible and your page is used, there is nothing to do but observe. If visibility stays low even though your page is retrieved, this is your fast lever: the existing page has to become more convincing. If even the retrieval is missing, only new content or authority from outside helps. So every question gets its own measure, derived from its situation.

Why this is an ongoing process

A one-off analysis is a photograph. Useful, but quickly out of date. Visibility in AI answers shifts constantly: through new content, through competitors, with every model update. What is a gap this week can be closed next week and torn open again somewhere else.

That is why I treat the analysis not as a project with an end date but as a recurring run. With a handful of questions you can do it by hand. With fifty or a hundred tracked questions, the sorting by retrieval against citation becomes the actual work, and that is exactly what can be automated. I captured the process in a small skill for Claude. It goes through each question one by one, pulls visibility, own URLs and third-party sources, and outputs a list of measures sorted by impact. Plus a table that drops straight into a spreadsheet. Started as a recurring run, say once a month, it puts the new fast levers on the table by itself each time, without anyone clicking through a hundred prompts manually. I have open-sourced the skill; it can be adapted to your own data source.

What you can do this week

You need no big setup for this. Take five questions where you want to be named in AI answers. If you have a tool for AI visibility, search it specifically for the “retrieved but not cited” pattern. Without a tool, ask the AI systems themselves and note for which of these questions you are not named at all, even though you have a fitting page. Those are your first suspects.

Take on the first such page and ask a single question: is there a clear, liftable answer to the customer’s question anywhere here, or does the text only circle it? Phrase that one answer, clean and verifiable, and watch over the next few weeks whether the citations pick up.

And put the analysis on a fixed rhythm. Higher frequency does not help on its own. It helps because the valuable gaps move. Whoever looks regularly always works on the levers with the best ratio of effort to impact.

Visibility here, too, is not an end in itself. A page that is retrieved but not cited is revenue already standing in the anteroom. Opening the door is usually less work than starting over.