Skip to main content
Prompt Research finds the best prompts buyers send AI assistants about a topic. Send a topic, and the result lists each product kind’s plain request, then the same request with the conditions buyers add, such as a budget, a skin type or an industry. Each prompt carries the monthly searches behind it, prompts[] is in recommended order, and brands[] shows which brands draw search demand in the same space. Use the prompts to plan content, ask them on answer engines or track the answers in a GEO monitor.

How it finds the prompts

  1. Product kinds. Keyword data for the market shows the kinds of product people search for under the topic, such as sunscreen stick under sunscreen. They are the result’s cores[].
  2. Conditions. For each kind, the research finds the questions that narrow a buyer’s choice, such as skin type, budget or team size. They are the result’s cores[].questions. Their options become conditions, together with the variants people search for (tinted sunscreen, sunscreen for oily skin).
  3. Prompts. Conditions combine into prompts up to depth, each with its measured or estimated searches, in recommended order.
A research takes about 1 to 3 minutes.

Submit a research task

The task is accepted and queued:
Read the result with GET /v1/async/task/{id}, or add webhook.url to the request and receive it when the task finishes. The task works like any other async task, and the result is in response. The request schema is in the API reference.
Submit Prompt Research through POST /v1/prompt-research only. POST /v1/async/task rejects taskType: "PROMPT_RESEARCH".

Request fields

The body is flat: the research fields and the task options sit at the top level.

Markets

  • US writes prompts in English. Demand is the monthly average over the last 12 months (demandSource.period: "monthly_average_last_12_months"), so a seasonal term reads lower at its peak than that month’s searches.
  • KR writes prompts in Korean. Demand counts the last 30 days (demandSource.period: "last_30_days"). The seed is matched against Korean search data, so send it as Korean searchers write it.
language in the result is en for US and ko for KR.

Depth

depth sets how many conditions one prompt combines, and every depth up to it appears in prompts: Options of one question are alternatives, so a prompt pairs oily skin with a budget and never with dry skin. Each depth needs at least one prompt, so limit must be at least depth + 1. depth: 0 returns each kind’s plain request and the questions people search about the topic.

Read the result

For example, a sunscreen run in the US returns a result like this, trimmed to a few entries:

prompts[]

The prompts in recommended order, up to limit entries. Each core’s prompts start with its plain request and its two price levels, then follow by demand across depths. The cores take turns in proportion to their searches, every plain request first, so any first part of the list is spread like the whole. Questions people ask to learn about the topic (intent: "informational") hold about a fifth of the list when the research finds them, because buyers ask them before they choose. To rank by demand alone, sort by monthlySearchVolume.

cores[]

The product kinds researched, the seed’s own first. Send cores in the request to research the kinds you choose.

brands[] and other fields

brands[] lists brands, manufacturers and product lines people search for in this space, with their monthlySearchVolume. Their demand stays out of prompts, because prompts do not name brands. Use them as the competitors to compare your brand with.

Score prompts against your brand

Send brand with brandDescription and each prompt gets a fit:
  • core: the description says your brand offers what the prompt asks for.
  • related: anything the description does not mention.
A product kind or condition the description rules out, such as another audience or product, is left out, so a result can hold fewer than limit prompts. fit ranks the prompts your brand answers higher; the demand numbers stay the same.
For example, the prompts in such a result carry fit like this:
The description is judged as written, so a one-line description leaves most prompts related. To keep a subtopic out of the result entirely, use exclude.

From research to a monitor

To track how often AI answers mention your brand on these prompts, feed prompts[] into POST /v1/monitors. This takes the finished research task, uses its prompts as the monitor’s prompts and the top ten brands[] as competitors (a brand monitor accepts up to 10), and creates a daily brand monitor:
The list is in recommended order, so the first N prompts make a balanced N-prompt monitor. A monitor’s run may queue at most your account’s tasksPerRun limit (200 on a self-serve account; GET /v1/monitors/capabilities reports yours), counted for each engine as prompts × location points × samples. Forty prompts on five engines with one point and one sample each is 200 tasks.
Creating a monitor starts its first run right away, and every run is charged like the tasks it enqueues. Check the prompt and engine counts before you create it.
For a category monitor, send "mode": "CATEGORY" with a category, and when the research has more than one core, map each prompt to its core in promptTopics. See Monitors for both kinds and the full request.

Reading the numbers

  • Depth 0 and 1 are measured. A plain request carries its kind’s searches, and a prompt with one condition the searches of its keyword (conditions[0].keyword).
  • Depth 2 and up are estimated. A prompt with two or more conditions is rarely searched as written, so its number is an estimate (volumeBasis: "estimated"). Use the estimates to rank prompts.
  • Broader prompts include narrower ones. The searches for sunscreen include those for mineral sunscreen, so do not add prompts’ numbers together.
  • Under 10 searches a month reads null. volumeReason says why: not_searched (nobody searches the condition), part_not_searched (one of its conditions is not searched) or estimate_below_floor (the estimate is under 10). Options from cores[].questions stay in the result even when nobody searches them, so a research on a business product often returns many null prompts.
  • The numbers are search demand. They count searches for the need behind each prompt in the market, over demandSource.period.

Errors and billing

A completed task costs 14 credits. A failed task is not charged. See Credits and limits.

Rejected at submit

Authentication, credit and concurrency errors are the same as for any task. See Errors.

Failed tasks

A failed task has status: "FAILED" and an error string that starts with its code:

Next steps

GEO monitors

Track the prompts you found across AI engines on a schedule.