Prompts & Running a Probe

Prompts are the questions your customers ask an assistant. They are the entire input to the measurement: the scores you get answer these questions and no others. A library of ten vague category prompts produces a precise-looking number about nothing in particular.

Building the library

Prompts tab. Three ways to fill it, and they mix freely.

Generate with AI

Choose how many prompts to produce per intent. The generator knows your brand and industry, and writes questions in shopper language rather than keyword language.

It works under hard rules: category and scenario prompts may not name any brand; comparison and trust prompts must name yours; and no prompt may name a country, because the market is applied at run time. See prompt intents.

Generated prompts arrive with their intent already known, since they’re requested and returned per intent rather than classified after the fact.

Batch import

Paste your own, one per line. Duplicates against the existing library are detected before import, and every imported prompt is classified into an intent automatically — with a deterministic fallback if classification is unavailable, so an import never fails halfway.

Good sources for imports: your search console queries, sales call questions, support tickets, and the questions your competitors’ content is written to answer.

From Discover

The Discover tab turns website analysis, Google Trends queries and Reddit threads into probe-ready prompts. Each one shows the signal it came from, and adds to the library in one click.

Reviewing the library

Each row shows the prompt text, its intent, and its source (AI, Reddit, GEO, Manual). Prompts that name your brand are flagged names brand.

Two habits worth keeping:

• Check the intent column after an import. Classification is good, not infallible, and the intent decides which rubric grades the answer and which averages the probe counts toward.
• Delete prompts you no longer care about rather than leaving them unselected. They still show up in reports and history as absent, and a smaller, sharper library is easier to reason about.

Running a probe

Select the prompts to run → Review & Run Probe. The dialog asks for three things.

1. Prompts Whatever you selected, editable from inside the dialog.
2. AI models Which assistants to ask. More engines is a wider read of the market; fewer is a cheaper, faster run.
3. Regions Which markets to probe. The dialog says whether your selected engines are region-aware at all — if none of them are, regions won't change the run.

The dialog then shows Total Operations — the exact number of probes, which is the number deducted from your allowance:

probes = prompts × Σ_engine (region-aware ? regions : 1) Region-agnostic engines (Claude, Doubao) run once in the primary market regardless of how many markets you selected.

Hit Run Probe and the job continues in the background. You can navigate away, close the tab, and come back to it.

One run at a time per target

While a run is in progress, starting another for the same target is blocked. Runs are the unit of history — overlapping them would put two different measurements on the same point in the trend.

What happens during a run

For each probe: query the engine → capture the answer and its citations → grade it against the rubric for that prompt’s intent → extract competitors and sources. Failed probes are retried up to three times within a fixed time budget; probes that still fail are marked not analyzed and excluded from every metric rather than scored as zero.

When every probe is done, run-level statistics are computed and an interpretation of the batch is generated: an objective reading plus prioritised actions, shown at the top of the visibility panel.

How often to run

Cadence	Fits
Weekly	Active GEO work — you want to see what a change did
Fortnightly	Steady-state monitoring for most brands
Monthly	Reporting rhythm, larger libraries, more markets

Answers move on their own between runs; assistants update their retrieval constantly. A single run is a snapshot, and a trend over several runs is the thing worth acting on. Keep the prompt selection, engines and markets stable between runs — changing the input changes what the trend means.

Reading the panel → What the run produced, tab by tab. Prompt intents → Why the mix of intents decides what you can learn.