How to Break Down Your AI Citation Share by Buyer Persona Instead of One Overall Number

Your AI visibility dashboard says 28 percent. It feels like progress, so you report it, and the number goes in the board deck next to a green arrow. What that single figure hides is that you are cited in 60 percent of answers to the questions your existing customers ask and almost never named when a brand-new buyer, the one you actually need to win, asks the question that starts their search. The aggregate is not wrong. It is just answering a question no one on your team is trying to act on.

This is the core weakness of a one-number visibility metric: it is a weighted average across audiences with opposite results, and averages hide exactly the variance you need to see. Segmenting citation share by buyer persona turns a flat score into a map of where you win, where you vanish, and which gap is worth closing first. Here is how to build that segmentation, how to read it, and how to convert a persona blind spot into a content plan.

Two metrics people mix up, and why the distinction matters here

Before you segment anything, separate the two numbers that get called "citation share," because segmenting the wrong one produces a tidy chart that means nothing.

The first is mention rate, also called AI visibility: the percentage of your tracked prompts where an AI answer names your brand at all. It is an absolute figure, as in "we appear in 22 percent of answers." The second is share of voice: your brand mentions divided by the total brand mentions across every company named in the same answer set, times 100, a relative figure that tells you how much of the conversation you own versus competitors. The canonical share-of-voice formula and the distinction from raw visibility are worth getting straight, because persona segmentation is useful on both but they answer different questions: mention rate tells you whether a persona can find you, share of voice tells you whether you beat rivals for that persona's attention.

Segment both. A persona where your mention rate is high but your share of voice is low means the engine names you alongside five competitors; a persona where mention rate is near zero means you are simply absent. Those need different fixes, and the aggregate number shows neither.

Why the aggregate number lies to you

AI visibility varies along three axes at once, and averaging across all of them is what produces the misleadingly stable overall figure.

It varies by engine. A brand can hold 40 percent of mentions in ChatGPT and 15 percent in Perplexity from the same web presence, because each engine reads a different index and weights sources differently. Blending them hides an engine where you are losing badly.

It varies by query intent. The same brand often leads on educational "what is" prompts and trails on "best tools for" comparison queries, because the two pull from different source types. A high score on definitional questions can mask total absence from the buying-intent questions that actually precede a purchase.

And it varies by persona. Different buyers ask different questions in different language, and an engine that names you for a technical evaluator's phrasing may never surface you for an executive's. Tooling that lets you segment visibility by persona and use case shows you where you win and where you vanish, which a single generic prompt set flattens into one deceptive average. When 73 percent of B2B buyers now use AI in their research, a persona you are invisible to is a segment of pipeline you cannot see leaking.

Build a persona-segmented prompt set

The segmentation is only as honest as the prompt set behind it, so this is where the real work is. Start from your actual personas, not demographic labels but the distinct buyers who ask distinct questions: the practitioner who evaluates, the director who scopes, the executive who approves, the adjacent role who influences.

For each persona, write the questions that person actually types into an AI assistant across their journey, in their own vocabulary. A practitioner asks "how do I" and "does X integrate with"; an executive asks "is this worth it" and "what do companies like us use." Aim for 100 to 200 prompts total to start, distributed across personas so each one has enough questions to produce a stable read rather than a noisy handful. Tag every prompt with two labels: its persona and its intent, whether educational, comparison, or buying. Those tags are what let you slice the results later.

Getting the questions right is the highest-leverage step, and it has its own discipline in how to design a prompt set that reflects how buyers actually ask AI about your category. Run each prompt across the engines your buyers use, several times, because answers vary run to run, and record whether your brand appears and in what position. Then aggregate up: mention rate and share of voice, computed separately for each persona and each intent tag. The mechanics of computing share across engines are covered in how to measure your brand's AI citation share across LLMs.

Read the persona matrix

Lay the results out as a matrix, personas down the side, and you will see the story the aggregate erased. A realistic example:

The aggregate of 28 percent looks like steady mid-tier visibility. The matrix shows something else: you are well known to people already using you and nearly absent to the executive and the influencer who decide and shape new deals. That is not a "publish more" problem, it is a "publish for these two personas" problem, and only the segmented view names it. The same logic scales into a board-ready format in how to build a board-ready AI visibility dashboard in a spreadsheet, where the persona rows are what turn a vanity number into a decision.

One caution on reading the matrix: set a noise threshold. Because AI answers shift run to run, a persona with only a handful of prompts can swing wildly between reads. Treat a low score as real only when it holds across repeated runs and enough prompts, so you chase genuine blind spots rather than sampling noise.

Intent segmentation belongs inside each persona

Persona is the first cut, but the sharpest insight usually comes from the second: query intent within a persona. A director might have a respectable overall mention rate that collapses when you filter to comparison queries, meaning you are cited when they are learning the category and dropped exactly when they are choosing between vendors. That is the most expensive place to be invisible, and it hides completely inside a persona-only view.

So slice each persona's prompts by the educational, comparison, and buying tags you attached earlier. The pattern to hunt for is a persona that is visible upstream and absent downstream, because that is a buyer the engine hands to a competitor at the moment of decision. Setting a realistic target for each cell, rather than one global goal, is what benchmarks and a target-setting playbook for AI citation share are for: a 9 percent share of voice with a scoping director is a different assignment than 9 percent with a practitioner, and the target should reflect it.

Turn a persona blind spot into a content plan

A segmented matrix is only worth building if it changes what you publish next. Read the matrix as a ranked to-do list: the highest-value cell is usually a persona with meaningful buying-stage query volume and near-zero share of voice, because that is a buyer actively deciding who cannot find you.

For that cell, the fix is content built in that persona's language answering that intent's questions, plus the entity and structured-data work that makes an engine confident enough to surface you for a new audience. Maintaining the machine-readable facts an engine draws on to name you for an unfamiliar persona is the job of the AI Feed Engine, and making sure a crawler reaches the pages written for that persona is a two-minute setup with a free llms.txt generator. The full loop, from segmented measurement to the content and structure that close a specific gap, is the system in how OnlyAEO works, and the pattern of moving from a blind spot to real presence is what the FastTrackr AI case study documents, where targeted work against specific buyer questions, not undifferentiated volume, is what moved the numbers.

Re-measure the same matrix on a fixed cadence so you can watch a single cell climb rather than staring at an aggregate that barely moves. The overall number will shift slowly even when a persona breakthrough is dramatic, because the average dilutes it. The persona row is where you see the win, and it is the row you show the buyer who asked for proof the work is landing.

The takeaway

One AI citation share number is a weighted average across engines, query intents, and buyer personas that get opposite results, and averaging them together hides the variance that should drive every decision. Split the number: build a persona-tagged, intent-tagged prompt set of 100 to 200 questions, compute mention rate and share of voice for each cell, and lay it out as a matrix. The aggregate that looked like steady progress will resolve into a clear picture of the buyer you already own and the one deciding right now who cannot find you. That segmented view is not a nicer chart. It is the difference between reporting a number and knowing what to write next.