Cluster 4 · Measurement

4.3 Prompt taxonomy, Fragility and Volatility

The quality of your GEO measurement stands or falls with your prompt set. A poor prompt set delivers figures that cannot carry a decision. This page describes how to categorise prompts — and introduces two Groundbase concepts for stability: Fragility and Volatility.

The four prompt categories

Direct brand questions

Prompts in which your brand name appears explicitly: ‘How would you rate X?’, ‘Is Y reliable?’. These measure the base knowledge and sentiment AI channels return about your brand.

Category questions

Prompts about your product category without a brand mention: ‘Best mortgage for first-time buyers’, ‘Which CRM suits consultants best?’. This is where you measure your real orientation position.

Problem-driven questions

Prompts from a user situation: ‘I’m buying my first house, where do I start?’. The type of question users actually arrive with.

Comparative questions

Prompts that explicitly ask for several players: ‘Compare X and Y for car insurance’. Important for the decision phase.

Prompt design principles

Fragility

Fragility — a Groundbase concept — describes how strongly your presence depends on the accidental phrasing of a question. If you appear for ‘which installer of automatic kitchen suppression systems in Flanders?’ but not for ‘who installs automatic suppression systems for commercial kitchens in Flanders?’, your visibility hangs on words, not on your position.

The pitfall is in what you call a “variant”. A real variant says the same thing in other words: same category, same market, same audience, same decision question. ‘Kitchen suppression systems for restaurants in Belgium’ is not a variant of the question above — different audience, different geography. A different answer to that is correct behaviour by the system, not fragility of your brand. Mixing those up means measuring noise and calling it insight.

Three to five clean variants per core question give a first indication: if you appear everywhere, your presence does not hang on one phrasing; if you appear only for one accidental wording, it is fragile. That is an indication, not a score — how Groundbase formally scores Fragility (number of variants, runs, weighting, uncertainty) is part of its own methodology.

Volatility

Volatility — likewise a Groundbase concept — describes how stable your presence is when identical prompts are repeated. AI channels are not deterministic: the same question can yield different answers at different moments.

Repeated measurement mixes two things you want to keep apart. Repeatability within one measurement window: several runs shortly after one another, with configuration held as equal as possible — this captures the system’s own chance variation. Drift over time: the same fixed set run again weeks or months later — where model updates, changed retrieval and changed facts come in. Throwing both on one heap means being unable to tell a channel’s product update from ordinary noise. So measure them separately, and log per measurement what was held fixed: channel, mode, language, location, date.

High volatility on core prompts is often a signal to take seriously: your visibility then depends more on chance than on structural presence.

One temptation to resist: explaining high volatility as “an open market where no brand is anchored yet”. It can be — it is a hypothesis worth testing — but volatility alone does not prove it; it can just as well come from retrieval, current data or the question itself. Look alongside it at competitor concentration and average presence: low volatility with one dominant competitor looks like an established market; high volatility with diffuse presence can be a market in motion. Treat that reading as a starting point for investigation, not a conclusion.

Using them together

Fragility tells you how robust your presence is across phrasings; Volatility how stable it is across repetition and time. The strongest position combines low Fragility with low Volatility: you appear regardless of the wording, and you keep appearing. Most worrying is the reverse — then your visibility is essentially noise.

verification
documented
That AI channels are non-deterministic is directly observable and described by the providers themselves.
own measurement
Fragility and Volatility are Groundbase concepts; the indications on this page come from our own measurement practice.
not verifiable
Whether variation stems from the model, retrieval or a product update — which is why we measure repeatability and drift separately.