Measure AI visibility for different buyer personas without mixing the results

To measure AI visibility for different buyer personas without mixing the results, track each persona as its own audience against the same set of prompts, keep an unbranded General baseline alongside them, and only compare rows that have the same prompts, engines, countries and time window. A persona result then tells you how the answer changed when the question was asked on behalf of that kind of buyer. It does not tell you what real people of that type see, because a persona is simulated context added to a question, not demographic data about actual users.

In short

• A persona is a named buyer profile added to the question your monitoring tool sends. Treat its results as "the answer when asked as this buyer", not as evidence about real users.
• Keep a General baseline for every prompt you give a persona. Without it you cannot tell a persona effect from ordinary variation.
• Compare only like populations: same prompts, same engines, same countries, same date range, similar answer counts.
• Small audiences are noisy. Mark thin rows as provisional and do not rank personas on them.
• Each persona you add to a prompt uses more prompt slots, so choose two or three personas that map to real buying roles.

What is a buyer persona in AI visibility tracking?

In AI visibility tracking, a persona is a named buyer profile, a short name and description such as "Procurement lead", that you attach to a tracked prompt so the question is asked on behalf of that kind of buyer. The plain version of the prompt is the General audience: the question as written, with no buyer profile.

In DiscoveredBy, a persona is not a separate prompt. It is an extra audience you attach to a prompt you already track (see Personas). A prompt is a question you ask AI engines on your behalf; it is not a keyword and not something a real user typed (see Prompts).

The mechanism is small. For a persona variant, the message sent to the engine gets an audience block appended after the prompt text, naming the persona, giving its description, and asking the engine to answer for that person. That is all that separates a persona run from a General run. It is not a logged-in consumer session, a saved profile the engine remembers, or a separate account.

Enlarge image

Why do persona results get mixed up?

Persona results get mixed up when rows that are not comparable end up in the same chart or sentence. Four causes are common, and each has a fix later in this post.

• Different prompts per persona. If the procurement persona tracks pricing questions and the compliance persona tracks security questions, a difference between them is a difference between questions.
• Different engines. Personas do not run everywhere. Google AI Overviews, Google AI Mode, ChatGPT (app) and Gemini (app) receive only the prompt text, so a persona variant is left out on all four rather than run as if it were General. A persona row never contains answers from those engines, while your General row may. Comparing a persona to a pooled General figure quietly compares different engine mixes. The engines reference and Explorer both spell this out.
• Different sample sizes. A persona with a handful of answers can swing widely from one window to the next. DiscoveredBy marks any audience with fewer than 30 analysed answers as provisional in the By persona views (see Sentiment trends).
• Different dates. Editing a persona changes what future runs send, while past runs keep what they actually sent. A trend that crosses an edit spans two different audiences under one name.

How many personas should you track?

Track two or three personas, each tied to a real buying role you can name in your sales process, plus General. A persona earns its place only if the role asks different questions or weighs different criteria than your default reader.

Every country, audience and language combination is its own prompt target and uses one prompt slot. One prompt tracked in 3 countries as General plus 2 personas, all as written, creates 3 x 3 x 1 = 9 targets and reserves 9 slots. How many active personas you can have depends on your plan (see Plans and limits).

So the practical rule is to apply personas to your most decision-relevant prompts, not the whole list. If you have not yet settled which prompts deserve that, read How to choose AI tracking prompts that reflect buying decisions and Audit your prompt list before adding more prompts first.

How do you build a persona prompt matrix?

Build the matrix by listing prompts down the side, audiences across the top, and marking which cells you will actually run. Then check that every persona column shares prompts with General.

Start from prompts you already track and trust. For each, decide whether the buyer role would plausibly change what the answer emphasises. A prompt like "best document review software for legal teams" might change for a compliance lead (audit trails) versus a procurement lead (contract terms). A prompt about a category definition rarely does.

DiscoveredBy can also propose questions per persona. Each persona card offers Suggest prompts, which asks a model for up to 10 buyer-intent questions that persona would plausibly ask, without naming your brand. They arrive as pending suggestions tagged with the persona, never tracked automatically, so you review before anything runs. Treat them as candidates, not a finished set.

Enlarge image

One caution: accepting a persona-tagged suggestion creates that persona's targets on top of whatever the prompt already runs, and General can be deselected on a prompt. Check that a General target exists for the same prompt, because the baseline is what makes the persona row interpretable. Plain CSV or Excel imports and onboarding only ever create General targets.

How do you compare persona results fairly?

Compare a persona to General on the same prompts, the same engines, the same countries and the same window, and read the difference as a description of that comparison, not a cause.

Apply this in order:

1. Restrict engines first. Because four engines never run persona variants, filter General to the engines the persona actually ran on before comparing. In Explorer, persona is a dimension, so you can break down or filter any metric by audience.
2. Match prompts. Use only prompts that have both a General and a persona target active in the window.
3. Check answer counts. Note the number of analysed answers per row. Under 30 is provisional, which is a sample-size flag, not a statistical confidence interval.
4. Read framing, not just presence. The By persona view on Sentiment and Brand reasons shows one row per audience with analysed answers, positive share, negative share, and, on plans with the entitlement, the top strength and top weakness reason (see Brand reasons).
5. Describe, do not explain. Write "when asked as a procurement lead, 6 of 18 analysed answers gave Quillstone's pricing as a weakness." Do not write "AI thinks procurement leads will dislike us".

Enlarge image

A stated reason is what an answer said. It is an observation of that answer, not proof of why a model produced it, and it does not reveal a model's internal ranking.

The audience worksheet

Copy this into a spreadsheet and fill one block per persona. Its purpose is to force the comparability checks before anyone draws a conclusion.

AUDIENCE WORKSHEET

Persona name:
Real buying role it stands for (job title or committee seat):
Where you know this role exists (sales calls, CRM, support tickets):
Description sent to engines (write it as a buyer would describe themselves):
Owner of this persona definition:
Date of last edit to the description (trends before this are a different audience):

PROMPTS
Prompts run for this persona (list ids):
Prompts also run as General (must be all of them):
Why this role would ask this differently:

POPULATION
Engines this persona ran on:
Engines excluded because personas are not sent there:
Countries and languages:
Date window compared:

COUNTS
Analysed answers, persona:
Analysed answers, General (same engines only):
Both at or above 30? If not, mark "provisional":

OBSERVATIONS (describe, do not explain)
Brand named in answers, persona vs General:
Positive / negative share, persona vs General:
Top stated strength and weakness, persona vs General:
Sources cited that differ:

CLAIM STRENGTH
Strong (large, comparable sample) / Tentative (provisional) / Anecdotal (single answers):

FOLLOW-UP
Hypothesis to test:
Page, source or question to review:
Re-check date:

Fill the "Claim strength" line honestly. Most persona findings will be tentative.

Worked example: Quillstone

Illustrative example: Quillstone and its competitors are fictional, and the numbers are made up to show the method.

Quillstone sells document-review software to mid-sized legal and compliance teams. It tracks one prompt, "best document review software for legal teams", in two countries, as General plus two personas: "Compliance manager" (audit trails, defensible review) and "Procurement lead" (contract terms, total cost). That creates 2 x 3 x 1 = 6 targets and reserves 6 slots.

Comparing over the same 28-day window, restricted to the engines that ran all three audiences:

The team writes three sentences and no more.

1. Compliance manager answers were more often positive than General (75% against 65%), on comparable sample sizes. That is worth watching but not yet a finding, because it is one window.
2. Procurement lead answers were positive in 9 of 18, but the row is provisional, so the team does not rank it below General.
3. In the stated reasons, "audit trail" appears as a strength in the compliance answers but not in General, and pricing is given as a weakness in 6 of the 18 procurement answers. The team logs both as hypotheses: the pages a procurement lead would read may under-answer pricing and contract questions. They plan to review those pages, not to declare that procurement buyers dislike Quillstone.

A follow-up for the procurement row: wait for more answers in the next window, keep the persona description unchanged, and re-read. If the count reaches 30 and the gap persists, it graduates from tentative to a finding worth acting on.

What persona tracking cannot tell you

Persona tracking has firm limits, and stating them is part of using it well.

• It is not demographic telemetry. The persona is text you wrote. It says nothing about what real procurement leads see in their own accounts, which engines they use, or how they phrase questions.
• It does not prove personalisation. Engines may vary an answer because of the audience block, or simply because answers vary from run to run. A difference between persona and General is an observed difference, not evidence that the engine models a buyer type. The General baseline and repeat windows are how you separate the two.
• It is engine-limited. Persona variants do not run on Google AI Overviews, Google AI Mode, ChatGPT (app) or Gemini (app), so nothing you learn from personas applies to those surfaces.
• Your description is a variable. Two different descriptions of "procurement lead" can give different answers. Record the description and the date of every edit.
• It multiplies cost and noise. Each audience uses prompt slots and splits your answers into smaller groups.

Persona wording is one kind of prompt edit. If you are deciding whether a change like "for a small team" alters the question itself, see Does adding "for a small team" change the question you are measuring?, which covers wording changes; this post covers the persona workflow.

Common mistakes

• Comparing a persona to a pooled General figure that includes engines the persona never ran on.
• Adding a persona to every prompt because you can, then being unable to read any of them.
• Editing the persona description mid-quarter and reading the trend as one continuous series.
• Ranking personas on provisional rows.
• Treating a stated weakness as a cause, or a persona difference as a market fact.
• Forgetting the General target on a persona-tagged suggestion.

Frequently asked questions

Do AI engines really answer differently for different buyer personas?

They can, but a measured difference does not prove an engine models buyer types. Answers also vary between runs. Use a General baseline and repeat windows so you can tell a persona difference from ordinary variation.

How many personas do I need?

Usually two or three, each matching a role that appears in your buying process. More personas means more prompt slots and smaller samples per audience. The number of active personas you can hold depends on your plan.

Can I track personas on Google AI Overviews or ChatGPT's app?

No. Google AI Overviews, Google AI Mode, ChatGPT (app) and Gemini (app) receive only the prompt text, so persona variants are not run there. The General variant of the same prompt still runs on them.

Should I change my persona description if the results look odd?

Change it only deliberately, and record the date. Edits affect future runs only, so a trend spanning an edit covers two different audiences. If you need a different profile, consider a new persona and keep the old one's history intact.

Are persona results the same as segmenting my real customers?

No. Personas are simulated context added to a question. Segmenting real customers needs your own customer data. Persona tracking answers a narrower question: how do answers to this question change when asked on behalf of this kind of buyer.

Next step: set up your first two personas

Start with the audience worksheet above, choose the two roles that matter most in your deals, and attach them to your five most decision-relevant prompts, keeping General on. Then read the By persona views on Sentiment and Brand reasons, and log what you see as hypotheses. DiscoveredBy supports personas in Settings → Personas, and the Personas docs cover limits and behaviour in full. To try it, sign in or create an account.