Branded prompts tell you what an engine says about you once it already knows who you are. Unbranded prompts tell you whether you exist in the consideration set at all. Only the second predicts pipeline.
Both are worth tracking. Reporting them as one number is what turns a useful pair of metrics into a flattering and useless one.
Almost every prompt set that looks healthy is branded-heavy, and almost every team that discovers this discovers it after a quarter of good-looking reports.
What is each type actually measuring?
Different stages, with different mechanisms behind them and different clocks.
| Branded prompt | Unbranded prompt | |
|---|---|---|
| Example | “Is Acme any good for mid-market teams?” | “Best tool for X when you have fewer than 50 staff” |
| Tests | Recall and accuracy | Discovery and shortlisting |
| Answers the question | Does the engine know us, and is it right? | Would a stranger ever reach us? |
| Typical result | High and stable | Low and volatile |
| Moves when | Your facts change, or a model generation lands | Third-party sources change, or a competitor publishes |
| Predicts pipeline | No | Yes |
The fourth row is the trap. Branded results are high because you named yourself in the question, which guarantees the engine has something to say. That number will look excellent forever and will never move with your market position.
Why does a branded-heavy set look so healthy?
Because it is measuring a question the engine cannot really fail. Ask about a company by name and a model with any memory of it will produce a plausible description.
Blend that with a handful of unbranded prompts and the average is dominated by the easy half. A set that is 70% branded reports a comfortable number while the part that matters sits near zero underneath it.
A blended visibility figure across both types is not a measurement of anything. It is a weighted average of one question you cannot fail and one you probably are failing.
This is the same reason a share-of-voice figure needs its denominator published, covered in how to calculate AI share of voice. If the composition of the set is not stated, the number cannot be compared to anything, including itself last month.
What do real unbranded prompts look like?
Longer and more situational than most tracked sets contain. Google’s May 2026 post on AI Mode reports that “the average AI Mode search is triple the length of a traditional Search query,” and notes growth in phrasings like “where to” and “where should I.”
That length is constraint. A buyer describing a situation names a team size, a budget, an integration or a regulation, and those constraints are what make the question answerable by a specific product rather than by the three market leaders.
- Shape
- Names you. “What does Acme do?”
- Use it for
- A regression check on accuracy. Report separately, never blended.
- Shape
- Names the category only. “Best CRM software.”
- Use it for
- Very little. Returns the same incumbents on every engine, permanently, so it produces a flat line.
- Shape
- Names the situation. “Best CRM for a two-person team that needs call logging under $50.”
- Use it for
- Everything. This is the winnable question and the one with revenue attached.
So what is a branded prompt good for?
Two real jobs, and both matter. Dropping branded prompts entirely is the overcorrection.
Accuracy. Asking what you charge, where you are based and what you integrate with is how you find out that an engine is confidently wrong about a checkable fact. That failure is silent, because a buyer told the wrong price simply disqualifies you without writing to ask.
Entity health. Consistent answers across engines mean your identity is resolved. Confident disagreement between engines on the same fact means two records exist somewhere, which is a different problem with a different fix, set out in what is an entity in AI search.
Track them, report them under their own heading, and never let them into the growth number.
What does it mean when branded is strong and unbranded is zero?
That you have a vocabulary or evidence problem rather than an awareness one, and it is the most common pattern in B2B.
The engine knows exactly who you are. It just does not file you under the thing buyers ask for, or has no third-party evidence connecting you to that category. Both are fixable and neither is fixed by publishing more about yourself, which is the argument in why doesnt ai put my product in the right category.
The diagnostic is cheap. Ask the branded question, confirm the engine is fluent about you, then ask the constrained unbranded question and see whether you appear. The gap between the two is the whole finding, and it points at third-party sources rather than at your own site.
How should the split actually look?
Weighted hard toward unbranded, with branded kept as a small fixed control.
Freeze the composition once you set it. Changing the branded share mid-quarter moves the blended figure without anything changing in the market, which is indistinguishable from a real trend on a chart.
Where do the unbranded prompts come from?
Not from a keyword tool, which is the shortcut that quietly reintroduces the problem. Volume data returns the compressed two-word forms people typed at a search box, and those are the unconstrained questions in the middle category above.
The three sources that produce genuinely constrained prompts are all records of things people said. Sales and support transcripts, where buyers state their constraints in their own words before your marketing has trained them out of it. Won and lost deal notes, which contain the situational half a keyword can never carry. And grounding queries from Bing Webmaster Tools, which are observed retrieval phrases rather than estimates of demand.
Twenty prompts pulled verbatim from transcripts will outperform a hundred generated from a volume tool, because the constraints are real. The test is simple: count the constraints across your set, and if the average is below one per prompt, the set came from a keyword tool whatever its stated provenance.
How many runs does each type need?
More for unbranded, and the reason is mechanical. OpenAI’s advanced usage guide notes that “Chat Completions are non-deterministic by default (which means model outputs may differ from request to request),” and variance is largest where the answer is genuinely contested.
Branded answers are stable, because the engine either knows you or does not. Unbranded answers about a contested category shift between runs, which is exactly where you need repeated sampling to get a rate rather than a coin flip. Budget the runs accordingly and state the count next to the figure, or the number cannot be defended by whoever presents it.