Branded vs unbranded prompts: why you need to track both

Two different questions hiding in one score

When people first measure AI visibility they usually load a list of prompts, run them, and report one percentage: how often the brand appears. That percentage mixes two situations that have almost nothing in common.

In the first, the buyer already knows you. They type your name, perhaps alongside a competitor's, and ask whether you fit their need. The engine is not deciding whether to mention you. It has to, because you are in the question. What you learn is how it describes you: accurate or not, positive or hedged, with which sources behind it.

In the second, the buyer has a problem and no brand in mind. The engine chooses which names to put in front of them. This is the moment where new customers are won or never find you. Here, being absent is the default, and you have to earn every mention.

What a branded prompt measures

A branded prompt names your brand, a product, or a person closely tied to you. Typical examples: 'Is Acme worth the price?', 'Acme vs Globex for agencies', 'Does Acme integrate with Slack?', 'Acme reviews'.

Visibility on these prompts is almost always high, because the brand is in the question. A score of 100% is normal and tells you very little. The useful signals sit underneath it: the sentiment of the claims the engine makes, which specific claims recur (pricing, support quality, accuracy, ease of use), which sources those claims lean on, and whether anything stated is simply wrong.

Think of branded prompts as brand defence. They tell you what a motivated buyer is told when they check you out before purchasing. A recurring doubt such as 'support is slow' or 'limited integrations' is worth more attention than any visibility percentage, because it is being repeated to people who are close to a decision.

Related reading: How to measure AI visibility (a practical method) and Share of voice in AI answers: how it is calculated and how to read it.

What an unbranded prompt measures

An unbranded prompt describes a need without naming you: 'best CRM for a small agency', 'how do I track AI visibility', 'cheapest way to run a coworking space in Bangalore'. These are the questions that decide who gets considered.

Here the metric is true visibility: out of the answers produced for this group of questions, in what share is your brand named at all, in which position among the named brands, and with what framing. It is normal for this number to be low and to vary by engine. It is also the number that moves when your content, citations and reputation improve.

Unbranded prompts should be grouped by the kind of question, because each group has a different job. Discovery questions ('best X for Y') build the shortlist. Comparison questions ('X vs Y') test you against named rivals. Problem questions ('how do I do Z') bring in people at the start of their research. A brand can win one group and be invisible in another, and the report should show that.

Why one blended number misleads

Suppose you track ten branded and ten unbranded prompts. Branded visibility is 100%. Unbranded visibility is 10%. The blended figure is 55%, which looks healthy and describes neither situation. Add five more branded prompts next month and the blended number rises without anything improving in the market.

The reverse also happens. A brand that dominates category questions but is described inaccurately when asked by name has a strong blended score and a real reputation problem.

Separating the two also makes targets honest. A sensible goal for branded prompts is accuracy and positive framing, not visibility, since visibility is already close to the ceiling. A sensible goal for unbranded prompts is share of the shortlist, measured against named competitors.

Setting up the split in practice

Start with a fixed prompt library, tagged at creation as branded or unbranded. Tagging by hand is less reliable than it sounds, because a prompt can mention a product, a founder or a misspelling of your name. The safest rule is automatic: if the text contains your brand name, an alias or a known product name, it is branded.

Aim for a mix that matches how people really search. Many teams find that branded prompts are a smaller share of the library than they expected, and that the larger part should be unbranded and comparison questions, since that is where growth is.

Report each segment with its own headline. For branded: percentage of claims that are positive, the top recurring doubts, and any factual errors found. For unbranded: visibility, average position when named, which competitors appear alongside you, and which sources the engines cite for those answers.

Reading the two segments together

The combination tells you what kind of problem you have. High branded and low unbranded: people who know you think well of you, but the engines do not bring you up unprompted. That points to content, citations and category association, which are covered in our guide to why AI cites competitors instead of you.

Low branded sentiment and low unbranded visibility: the engines know little about you and what they know is unflattering. Start with accuracy and the sources that shape claims about you before investing in reach.

High unbranded and low branded sentiment: you are being recommended to people who then see doubts when they check. Look at the specific recurring claims and answer them on your own pages, because engines will quote a clear, honest answer from a source they trust.

How OClever reports the split

OClever tags every prompt as branded or unbranded automatically and keeps the two segments apart through the whole product. The report page has a section for each: branded questions with positive sentiment, what engines praise and what they doubt (with the actual sentences they used), and unbranded questions grouped by type with average visibility per group.

You can also filter any dashboard view to one segment, so a visibility chart is never accidentally a mixture. If you want to test a single question before adding it to the library, Prompt Lab runs it live across engines and shows whether the answer names you.