Answer Radar / Prompt research
Propose the buying questions, approve the ones worth measuring
You cannot see what people ask ChatGPT. You can reason about it from your domain, your competitors, your category and what buyers type into Google. Answer Radar proposes a draft set from exactly that, and measures nothing until you approve it.
Updated 2026-09-04
What prompt research is here
A draft you approve, not a list that becomes your spend.
Pressing Generate questions, or calling answer_prompts_propose, asks Answer Radar to write commercial-intent buying questions from the project's domain, brand name, named competitors, category and buyer segment. They arrive as a draft set, every one unapproved. You keep or cut, assign a type and topics if you want them, and set priority. Only approved questions are measured, and where Search Console is connected the real query list is the check the proposal cannot be.
Our own prompt set, read on 4 September 2026
The propose call is shown for its shape and was not executed, because it writes a draft set to our production project. The list call is real.
answer_prompts_propose({
project_id: "<your project id>",
count: 8,
idempotency_key: "propose-2026-09-04"
})
answer_prompts_list({
project_id: "<your project id>",
status: "approved"
})Through the Linkeddit MCP connector. count is 1 to 20; the default is 8.
answer_prompts_list, summarised from the response
- prompts
- 24 in one prompt set, version 1
- status
- 21 approved, 3 removed, 0 proposed
- type
- null on all 24. Nobody has classified the set, and the tool did not do it for us.
- priority
- 86 to 100 on the fifteen unbranded category questions (best AI brand monitoring tools at 100, best competitive intelligence tools for B2B SaaS at 96); 1 to 10 on the Linkeddit-named questions.
- examples, approved
- What are the best Klue alternatives? Which competitive intelligence tools use Reddit and review site data? How can I find out why ChatGPT recommends my competitor instead of me?
- examples, removed
- Three Linkeddit-named plan and DIY comparison questions, removed rather than deleted so the history stays.
The nine open recommendation gaps on this project all came from the unbranded questions at the top of the priority order. The branded ones at the bottom measure whether the engines know us, which is a different question.
The six prompt types, with an example from our own set
Types are labels you assign. The examples are real prompts from the linkeddit.com project, except review, which our set does not carry.
| Type | Example |
|---|---|
| best_of | What are the best AI citation tracking tools? |
| alternatives | What are the best Klue alternatives? |
| comparison | How does Linkeddit compare to Apollo, ZoomInfo and LinkedIn Sales Navigator? |
| pricing | How much does Linkeddit cost, and what is included in each tier? |
| how_to | How do I monitor my brand's visibility in AI search results? |
| review | A question asking what a named product is like to use. |
Three refusals, and why each one matters
A proposer that did any of these would be more impressive in a demo and worse for your credit balance.
It approves nothing
Deciding what gets measured is a separate act from suggesting candidates. Every proposed question lands unapproved in a draft set, and the set stays draft until you approve one. A generated list never quietly becomes your spend.
It classifies nothing
Type, one of best_of, comparison, alternatives, pricing, how_to or review, and intent such as bottom of funnel are labels you assign after reading the questions. On our own project every one of 24 prompts still has type null, which is the honest state of a set nobody has classified.
It estimates no volume
No engine publishes how often a private assistant is asked a question, and the proposer will not invent a number to rank suggestions by. The check against demand is Search Console, where connected.
What the proposer refuses to do
Three refusals, each of which keeps a suggestion from becoming a decision.
- It does not approve. Approval is a human act, in the app or through answer_prompts_update, and the set stays draft until then.
- It does not classify. Type and intent are labels you assign after reading the questions.
- It does not estimate prompt volume. No engine publishes one, so none is printed.
- It does not touch approved questions when you regenerate. A new draft set is another set.
Frequently asked questions
How do I choose which prompts to track for AI visibility?+
Start from commercial intent: the questions a buyer asks an assistant before choosing a product in your category, such as best tools for the job, alternatives to a named competitor, comparisons and pricing. Answer Radar proposes a draft set from your project's domain, brand, competitors, category and buyer segment; you approve the ones worth measuring, cut the rest, and where Search Console is connected you check the set against what buyers actually type. Ten to twenty-five approved questions is a workable set.
Where do the proposed questions come from?+
From the project's commercial context: domain, brand name, the competitors you named, category and buyer segment. The proposer writes questions a buyer with commercial intent would ask. Through the connector it is answer_prompts_propose with a project id and a count from 1 to 20. Regenerate produces another draft set; nothing already approved is touched.
Why does this page not show a live proposal?+
Because proposing writes a draft prompt set onto the project, and we do not decorate a page by writing to our production account. The list call is read-only and its result is real: 24 prompts in one set on our linkeddit.com project, 21 approved and 3 removed, on 4 September 2026.
What are topics?+
Your own grouping of questions. Topics and tags are one mechanism and a prompt may carry several. They are stored lowercase with whitespace collapsed, so Pricing and pricing are the same topic. Metrics can be restricted to the prompts carrying one topic, and the response says how many prompts the filter matched so a filtered number is never read as an unfiltered one.
Should I track prompts that name my own brand?+
Sparingly. A question that names you measures whether the engine knows you, not whether it recommends you to a buyer who has not heard of you. Our own set carries a handful of Linkeddit-named questions at low priority and fifteen unbranded category questions at the top. The unbranded ones are where the nine open recommendation gaps came from.
Can my assistant do the research?+
Yes, and it is a better fit than a template. The connector exposes propose, list and update. An assistant can read your Search Console queries, read the existing set, propose candidates, and hand you the draft to approve, which stays a human step.
Read next
Search Console grounding
The first-party query list a proposed set should be checked against before you approve it.
Read HelpHow prompts are proposed
What Generate questions draws on, what it refuses, and the two ways to check a set.
Read GuideYou cannot see what people ask ChatGPT
Why buyer prompts have to be reasoned about, and how to do it without inventing data.
ReadMeasure the questions buyers ask, not the ones you hoped they ask
Prompt research is part of Answer Radar, included with the Compete plan, in the browser and through the MCP connector.