Do Brand-Name Prompts Inflate AI Visibility Scores?
Yes, brand-name prompts inflate AI visibility scores. A prompt that names your brand, such as "Is [brand] good for small teams?", almost always produces an answer that mentions you. Counting it alongside unbranded prompts makes your mention rate look far higher than a buyer's real chance of seeing you. Prompts that frame your brand favourably add a second distortion. Assistants tend to go along with the user's framing, a behaviour known as sycophancy, so they may endorse or include a brand simply because the question suggested it. Keep branded prompts in a separate group, and use unbranded "blind" prompts for mention rate and share of voice. Compare the two to see how much of your apparent visibility depends on being named first.
How much inflation do branded prompts cause?
A simple example shows the effect. Take a set of 100 prompts, 30 of them branded:
- On branded prompts, the brand is mentioned in nearly every answer: say 100%.
- On unbranded prompts, it is mentioned in 15% of answers.
- The blended mention rate is 30 × 100% + 70 × 15%, divided by 100, which is 40.5%.
The blended figure is almost three times the rate a buyer asking an unbranded question would experience. The more branded prompts a tool or agency includes, the higher the headline number climbs, without any change in real visibility. Check the mix in any score you are shown, including in case studies.
Should you track branded or non-branded queries for AI visibility?
Track both, for different purposes:
| Query type | Example | What it measures |
|---|---|---|
| Unbranded | "Best payroll software for restaurants" | Whether you are chosen when buyers do not name you: discovery |
| Competitor-branded | "Alternatives to [competitor]" | Whether you appear when buyers consider a rival |
| Own-branded | "Is [brand] good for restaurants?", "[brand] pricing" | Accuracy and sentiment of what assistants say about you |
| Head-to-head | "[brand] vs [competitor]" | How you are positioned in direct comparisons |
Only unbranded and competitor-branded prompts should feed mention rate and share of voice. Own-branded and head-to-head prompts should be reported separately, on accuracy, sentiment and framing.
How to compare blind and brand-anchored AI queries to spot sycophantic mentions
Assistants trained with human feedback tend to agree with the person asking. Academic research on this, including a widely cited 2023 study, found sycophancy to be a general behaviour of these assistants. In 2025, OpenAI rolled back a ChatGPT update after it made the model noticeably more sycophantic. For visibility measurement, the risk is that an assistant says "Yes, [brand] is a great option" because the question implied it.
To test for it, run matched pairs:
- Blind prompt: "What are the best payroll tools for restaurants?"
- Anchored prompt: "Is [brand] one of the best payroll tools for restaurants?"
- Competitor-anchored prompt: the same question naming a competitor.
- Negative anchor: "Is [brand] a poor choice for restaurants?"
Run each several times and compare:
- Anchoring lift: endorsement rate on the anchored prompt minus mention rate on the blind prompt. A large lift means the answer follows the framing rather than independent judgement.
- Symmetry: if the assistant endorses you and every competitor when each is named, the endorsement carries little information.
- Negative anchor response: if it also agrees you are a poor choice when asked that way, the answers are tracking the question, not the evidence.
A brand with a small anchoring lift and a solid blind mention rate has visibility that does not depend on being named first.
Do branded prompts belong in an AI share of voice report?
No, not in the share itself. Share of voice compares how often each brand appears when buyers ask neutral questions. Branded prompts guarantee the named brand a mention, so they tilt the share towards whoever the prompts name. Keep them out of both the numerator and the denominator.
Two exceptions are worth handling carefully:
- Competitor-branded prompts, such as "alternatives to [competitor]", are neutral from your point of view and can stay in, as long as the competitor's own mentions in those answers are excluded or reported separately.
- Category prompts that happen to contain a brand-like word should be checked, so a generic term is not mistaken for a brand.
Report branded results in their own section: accuracy, sentiment and framing. The KPI choices are covered in which AI visibility KPIs brands should track.
What other things turn blind prompts into branded ones?
- Chat memory and history. If the account has discussed your brand before, "unbranded" prompts may produce answers shaped by that history. Use logged-out sessions or clean accounts with memory off.
- Custom instructions on the testing account that mention your company.
- Follow-up questions in the same conversation after the brand has come up.
- Location or profile signals that point to your business, such as testing from your office's location for local questions.
Record the account state for every run, so these effects can be ruled out.
How should agencies present this to clients?
Show unbranded mention rate as the headline. Show branded prompts separately with accuracy and sentiment, and include a short note explaining the difference. If a previous report blended them, restate the history with the split, even if the headline falls. Clients trust a lower, honest number more than a high one that collapses under scrutiny. Checking published case studies for this problem is covered in how to detect cherry-picked prompts in an AI search case study.
What might an illustrative restatement look like?
Consider an illustrative agency that has reported a client's "AI visibility score" of 52% for six months. Reviewing the prompt set, it finds that 35 of 90 prompts name the client, including questions like "Is [client] good for enterprises?". It splits the set. Unbranded mention rate turns out to be 22%, and share of voice among five tracked competitors falls from first to third. Matched-pair tests show a large anchoring lift in one engine. There, the client is endorsed in almost every anchored answer but named in only one in six blind answers. The agency restates six months of history with the split, and explains the change in a short note. It then refocuses the work on the blind questions where the client is missing. The client's leadership accepts the lower figure, because it now matches what sales hears from prospects.
What do people ask about branded prompts?
Are branded prompts useless?
No. They are the best way to check what assistants say about you: accuracy, pricing, sentiment and positioning.
What share of a prompt set should be branded?
Enough to cover the accuracy checks you need, usually a small minority. Keep them in their own group regardless of size.
Do all engines show the same anchoring effect?
No. It varies by engine and model version, which is why matched-pair tests should be repeated after model updates.
Does sycophancy affect unbranded prompts?
Less, because the question does not suggest an answer. But follow-up questions in the same chat can reintroduce it, so run blind prompts in fresh sessions.
How can we tell if a vendor's score includes branded prompts?
Ask for the prompt list and the branded share. If they cannot provide it, treat the score with caution.
Published by The Kovmere Team · updated Sep 29, 2026 · More resources