Prompts
The questions your customers ask AI, run per engine when you choose.
Prompts are the unit of GEO tracking: the questions your customers ask AI answer engines, each run against a chosen set of engines when you start it. Every metric on the AI Visibility pages derives from prompt runs.
Anatomy of a prompt
| Field | Meaning |
|---|---|
| text | The question, 3–300 characters. Write it the way a customer would actually ask. |
| category | Free-form grouping (e.g. comparison, pricing, how-to). |
| engines | Which engines run it. Default: ChatGPT, Gemini, Perplexity; projects in China (cn / zh) default to Qwen, DeepSeek, Doubao. |
| active | Off pauses it without losing history. |
Writing good prompts
- Ask like a customer, not a marketer. "best managed postgres for startups" beats "why is Acme great".
- Include comparison terms. "acmecloud vs hetzner" is where purchase decisions happen.
- Cover the buying journey: awareness ("what is X"), comparison ("X vs Y"), selection ("is X good for Z").
- Don't stuff your brand into every prompt. Unprompted mentions (the engine brings you up on a generic question) are the strongest signal you can get.
Running prompts
- GEO runs only when you start it. There is no schedule.
- Run on a prompt opens a confirmation: pick the engines, see "about $x" and the maximum that will be held, and for DeepSeek choose off-peak (default) or now.
- Every
prompt × enginerun is billed at its actual cost (see Usage & balance). A re-run is a new run. A run the provider did not charge for costs nothing; a run that fails after the provider charged us is returned as platform credit. Failed runs are never retried automatically. - GEO calls paid third-party APIs, so it needs an active subscription. One request holds at most 50 runs, and an organization at most 2,000 a day.
- Scraped engines finish asynchronously; DeepSeek off-peak runs wait until 22:00 Beijing time.
What a run records
Each run stores the engine's full answer, the brands it mentions, the domains it cites (and, for Qwen and DeepSeek, the pages its search returned), the searches it ran when the engine exposes them, and how the answer was obtained (real product or official API — see Data sources). Mention detection matches your project's brand names and aliases against the answer text, which is why aliases matter (see Competitors).
Managing the set
Prompts are per project. A healthy set is 10–30 prompts across categories, reviewed monthly: retire questions nobody asks and add newly popular phrasings. Re-run volatile prompts (pricing, comparisons, new entrants) more often than stable ones.