Prompts

Prompts are the questions Surfaced asks the AI engines on your behalf. Every number you will ever read is an aggregate over them, which makes the prompt set the measurement instrument rather than a setting. Add and edit them on Insights → Prompts.

What makes a good prompt

A question a buyer would type. Not a keyword. “best AI visibility tracker for agencies” is a prompt; “AI visibility” is not, and an assistant asked the latter will write an encyclopedia entry nobody is in. A question with a decision behind it. The questions worth tracking are the ones where an assistant’s answer changes what someone buys: comparisons, “best X for Y”, “how do I choose”, “is X worth it”, “alternatives to Y”. A question you could lose. If you are the only possible answer, tracking it tells you nothing. The useful prompts are the contested ones. Unbranded, mostly. A question containing your name will return your name. See below.

Branded and unbranded

A branded prompt names you: “is Acme any good?”. An unbranded prompt does not: “best tool for X”. They answer different questions. Unbranded prompts tell you whether you are discovered. Branded prompts tell you how you are described once someone already knows you, which is useful and a different thing. Keep both, but never read them blended. Surfaced defaults to unbranded for this reason, and the gap between the two is routinely enormous: the same workspace can be near 80% branded and under 1% unbranded. More.

Stage, intent and tags

Each prompt carries a funnel stage and a decision intent, and any tags you give it. These are how you slice results later. Tagging by product line or by buyer type turns one visibility number into five you can act on. Decide a tagging scheme before you add thirty prompts. Retrofitting tags is dull work.

Languages and markets

A prompt is tracked per language and per market. Each combination is a separate row with its own numbers and its own response cost. See Markets and languages before you multiply.

How many

There is no limit on how many prompts you may save. What your plan meters is tracked responses: one prompt, asked of one model, in one language, from one market, once. The real constraint is therefore prompts multiplied by languages, markets and models. The toolbar on Prompts shows that arithmetic against what is left this month. So the question to ask is which prompts are worth running every day. For most businesses that is a few dozen. Resist completeness. Prompts covering a category you do not serve lower your brand visibility without telling you anything you can act on, and they consume the budget that would let you watch a real question in three more markets. Start with what your sales team gets asked, and add a prompt each time a real buyer asks something you were not tracking.

Tags are topics

The labels you put on a prompt are the same thing the product elsewhere calls topics. Some are inferred for you from the wording (comparison, features, pricing, use cases, alternatives, reviews, technical, general) and you can set your own on top. Tag edits apply to the whole group, including any inactive siblings.

Editing and retiring

Editing a prompt inside a group updates it in place. Editing a standalone prompt resets its analysis, and the toast says so. Either way, past answers stay attached to what produced them. Untracking is the safe way to stop a question running: the row and its history are kept, and an “Untracked prompts” section offers to track it again. Deleting is permanent and takes every answer with it. Prefer untracking unless you want the history gone.