Back to Insights
Guide October 30, 2026 Bora Kurum

How to Build an Unbranded Question Set

AI visibility is measured with questions that never name your brand. How the set is built, which questions stay out, and why the set must be frozen.

MethodMeasurementQuestion SetBaseline
In this article +

AI Quick Summary: An unbranded question set measures AI visibility with questions that never name your brand: 20 questions across 4 layers (category discovery, problem-led, definitional, branded control), written in real buyer language, frozen once measurement starts, and re-run in 30-60-90 day rounds.

The most common mistake in visibility measurement is testing the brand with its own name. Ask "does my brand appear on ChatGPT?" and type your name — the answer only shows the model recognizes that name. The commercially meaningful question is different: when a prospect describes the category, does your brand enter the candidate set?

This article explains how the question set behind measurement gets built.

The Set Has Four Layers

Not all questions do the same job. A mixed list hides which signal is missing. The split is built as follows:

Category discovery. Questions like "which GEO agencies operate in the UK?" that ask the category directly and expect names. The hardest layer to win, the highest in value: whether the brand sits in the category set gets decided here.

Problem-led questions. The user has a problem but doesn't know the category's name: "my brand never appears in AI answers, what should I do?" Retrieval on this layer shows the content genuinely answers.

Definitional questions. "What is GEO?", "How is AI visibility measured?" Lowest commercial value, fastest to win. Early signals come from here: whether passages are indexable shows up first.

Branded control. Not a KPI — a diagnostic. If nothing comes back when the brand name is asked, the problem isn't passage quality but entity recognition — and that needs entirely different work.

Layer distribution starts equal (5 questions each); as the category matures, weight shifts to discovery and problem layers. The definitional layer serves as early warning, branded control as diagnosis — neither is a KPI.

Question Writing Rules

Every question follows five rules:

  1. Real language. The sentence the sales team hears gets written verbatim; no lab sentences.
  2. Single intent. One question asks one thing; "what is X and which is best?" is two questions — split it.
  3. Unbranded. No brand name passes (except the branded-control layer); the category and the problem get described.
  4. No supremacy. Ranking invitations like "best" don't get asked; the model ranks, the score plays.
  5. Repeatable. The same question must be askable verbatim three months later; time-embedded phrasing ("this month") stays out.

A question breaking the rules doesn't enter the set; it gets fixed in draft. Question quality is measurement quality.

Sample 20-Question Skeleton

Adapt to your category; the skeleton stays fixed:

Category discovery (5): "Which X agencies operate in the UK?", "Firms offering X services?", "Who should I work with for X?", "Known brands in the X sector?", "My options for an X need?"

Problem-led (5): "My brand never appears in AI answers, what should I do?", "What matters when choosing X?", "Common mistakes in X?", "How is an X budget set?", "How long do X results take?"

Definitional (5): "What is X?", "How does X work?", "X vs Y difference?", "How is X measured?", "Who is X for?"

Branded control (5): "What is [Brand]?", "What does [Brand] do?", "Is [Brand] reliable?", "[Brand] pricing?", "[Brand] reviews?"

When filling the skeleton, put your category for X and your name for [Brand]. Read the questions to three real users; rewrite any question that fails the "would you ask it like this?" test.

The Expected Starting Picture

Near-zero scores on the category-discovery layer in the first measurement are normal. Read it as the measurement's floor, not a failure. The shape to watch is this: branded-control questions scoring high while category questions score zero. That shows the brand is recognized but not associated with the category — the most common picture of all.

Read the picture in four cells: branded high + category high (mature), branded high + category low (recognized but unassociated — most common), branded low + category low (ghost), branded low + category high (rare; may be a measurement error). Your cell sets your strategy.

Which Questions Stay Out

Supremacy-bearing questions don't enter the set: a question like "which is the best GEO agency?" invites the model to rank. Two consequences follow. First, ranking-inviting questions swing markedly from run to run and scores stop being comparable. Second, the desire to appear on these questions leaks into content over time and leads the brand to write unfounded supremacy claims about itself.

Likewise, factual questions with a single correct answer don't suit measurement. No visibility gets won on them; only the correct answer's source gets shown.

Third exclusion: current-events questions ("this week's X development"). Time-stamped questions can't compare across rounds; they don't enter trend measurement. Freshness gets tracked on a separate panel.

Counting Rules: How Is an Answer Scored?

Every answer gets scored with the same rules; the rules live in writing in the round file:

  • Mentioned: the brand name passes in the answer. The first 1-point threshold.
  • Cited: the brand is marked as a source. This is the real KPI.
  • Correct description: the brand passes in the right category. Wrong category scores half.
  • Competitor position: which competitors pass, where do you rank? Relative position gets noted.

Rule changes happen outside rounds and get filed; rules never change mid-round. Counting-rule details are in our citation article.

Every Answer's Raw Text Gets Kept

Logging only the score isn't enough. Seeing a rate fall doesn't say why it fell. When the raw answer text is kept, these become visible: who gets recommended instead of your brand, which phrases describe your brand, and which source the answer rests on.

One thing deserves special watching here: which category describes your brand. A platform company constantly called an "agency" in answers isn't a score problem — it's entity-definition drift, and it needs fixing on the content side.

The raw-text archive serves three jobs: finding drop causes, detecting category drift, watching competitor moves. A panel without an archive is measurement without memory.

The Set Gets Frozen

Once measurement starts, the question set never changes. When a question is added or removed, the next measurement stops comparing with the previous one and the program's entire progress record loses meaning. When new questions are needed, they open as a separate set and get tracked from their own floor.

The same rule covers the platform list and the language. The source pool pulled for a question asked in Turkish can differ from the English-asked one; two languages' results never merge into one average.

Freezing discipline: the set file gets versioned (v1, v2...); every round logs which version ran. Version changes happen only by opening a new set — the existing set is never edited.

When Measurement Happens

The baseline measurement gets taken before any visibility-affecting work starts. A baseline taken after press and external publishing began is no longer a baseline: what contributed how much can never be separated again.

Afterwards the same set re-runs at regular intervals. What matters more than the interval itself is running under identical conditions every time.

Baseline timing is critical: before the website revamp, before press work starts, before the content sprint. If the baseline slipped, the nearest "clean" date counts as the start and gets noted in the report — an honest baseline beats a late one.

Growing the Set: From 20 to 60

Frozen doesn't mean the program never grows; growth happens through new sets:

Set 2 — competitor questions. Comparisons naming competitors ("X or Y?"). High intent, narrow scope.

Set 3 — long tail. Niche sub-questions, sector-specific variants. Low volume, fast wins.

Set 4 — new language/market. An English panel or a new geography. Separate set, separate floor, separate trend.

Each set freezes with its own version; no averaging across sets. The report carries a separate curve per set. All four sets complete by the fourth quarter — don't rush; never move to the second before settling the first.

4 Common Mistakes

Starting with a branded question. The "does my brand appear?" test demoralizes or gives false confidence. The set is built unbranded; branded questions live only in the control layer.

Stealing questions. Questions copied from a competitor's FAQ don't reflect real user language. Questions get gathered from the field (sales, support, community).

Keeping the set alive. The sentence "let's add a question" kills the trend. Addition requests turn into new-set proposals; the current set stays frozen.

Throwing away raw text. The score table is kept, the answers deleted. Three months later the drop cause can't be investigated. The raw-text archive is mandatory — storage is cheap, memory is expensive.

Next Step

Build your set this week: write 20 questions, run them through the 5 rules, read them to 3 users, freeze. Take your first measurement with a Citation Baseline; the method is in our methodology, concepts in the glossary, evidence in the case study, general questions on the FAQ page. The set is your program's foundation — a solid foundation raises the building. Lay it this week; in three months the first floor completes and the curve speaks in your favor every round.

FAQ

Why is an unbranded question set necessary?

Testing with your brand name only shows recognition; the commercially meaningful question is whether you enter the candidate set when the customer describes the category. The set measures that question.

How many questions should the set hold?

20 questions suffice to start: 5 from each layer. Grow to 40–60 as scope expands; the set is frozen and new questions are tracked in a separate set.

When is the question set changed?

Never. Once measurement starts, the set is frozen; changes destroy the trend. New questions open as a separate set and are tracked from their own baseline.

About the author

Bora Kurum is the founder of RAG Signal and a Ph.D. researcher at Istanbul Bilgi University, where his work focuses on LLM retrieval behavior and brand representation in generative AI systems. Read more →

Ready to get your brand cited?

Start with a baseline AI Visibility Audit. Know where you stand across every major AI model.

Get Your Citation Baseline →

0% → 81% cited in 90 days · Real client measurement

Get Your Citation Baseline →