AI prompt research
Finding the questions people ask AI assistants about your market, what the demand estimate is and is not, and how a question becomes a tracked prompt.
AI prompt research answers the question prompt tracking cannot: which prompts are worth tracking in the first place. Give it up to eight seed topics and it returns the questions people ask AI assistants around them, in the market and language you choose, each with an estimate of the demand behind it and a read on whether you already measure something like it.
It costs 35 credits a run, on every plan, and a run that finds nothing usable costs nothing at all.
What the demand figure is, and what it is not
Read this before you read anything else on this page.
Every candidate carries a figure RankX AI calls estimated question demand. It is an estimate of how often the question is asked, modelled by our data provider from Google's People Also Ask results. It is not a count of ChatGPT or Gemini queries. No product can produce that number, because none of those platforms publishes it to anybody.
The figure is still worth having, for one specific job: it separates a question people demonstrably ask in public from one somebody invented in a meeting. Use it to order candidates against each other, not to forecast traffic.
The scale is smaller than search volume, by design. These are questions, not keywords. A broad market seed tops out in the low hundreds a month and a specific one in the tens, so a question with a demand of 20 is a strong result rather than a weak one. Reading these numbers against search volumes is the most common misreading and it makes every result look like a failure.
The three states of a blank column
RankX AI keeps three things apart that most tools collapse into one dash, because they lead to different decisions.
| What you see | What it means | Does it sort? |
|---|---|---|
| A number | The provider modelled a figure for that question, in that market. A modelled zero is one of these | Yes |
| No estimate | The provider has nothing for this question. That is not the same as nobody asking it | No, it sorts last |
| Could not be checked | The demand lookup itself failed on that run. That is our fault, not the data's | No |
A question with no estimate has its score marked partial rather than quietly topped up, so you can tell a complete score from an incomplete one at a glance.
Running a search
Seeds. Up to eight topics a run. Type them, or leave the field empty and RankX AI derives them from what your project already knows about itself: your services, your topic clusters, keywords linked to those clusters, your own Search Console queries, and searches assistants ran themselves while answering your existing prompts. Specific beats generic. A named service produces better questions than an industry label, and a one-word seed produces the broadest and least useful set of all.
Your own brand name is never used as a seed, and that is deliberate rather than an oversight. The corpus behind this feature is built from public search questions, so seeding it with a brand name returns almost nothing for any business that is not a household name. Brand-shaped prompts come from your Brand Book instead, which knows facts about your company that no search corpus does.
Market and language. Pick the country the questions should come from and the language they should be in. The list covers 94 markets, each with its own set of languages. If your project's market is not one of them, RankX AI says so and names the substitute it used rather than quietly running your research against the wrong country.
Depth. 25, 50 or 100 questions a seed. Start at 25 to find out whether a topic has anything behind it at all, then go deeper on the ones that do. It is a fixed menu rather than a free number so that a page cached at one depth can serve the next request.
Filters. Two, both optional:
- A minimum demand, off by default or 1, 10 or 50 a month. Off is the default for a reason: on a specific topic a floor can return nothing, and no questions is a worse answer than questions with no demand.
- A word it must contain, up to 60 characters, for narrowing a broad seed to one part of it.
How a candidate is scored
The score is 0 to 100 and it is not a demand ranking. It is built from four signals:
| Signal | Weight | What it reads |
|---|---|---|
| Relevance | 40 | How close the question is to what your business actually does |
| Demand | 25 | The estimate above, banded rather than used raw |
| Shape coverage | 20 | Whether you track anything of that question's kind yet |
| Source quality | 15 | Where the question came from. A question an assistant generated itself is stronger evidence than a modelled one |
Relevance carries the most weight because demand is the weakest of the four on a specific topic, where most questions legitimately have no figure. A rare question about something you actually do outranks a common one that is not about you, and that is the behaviour you want from a research tool.
Question kinds, and why coverage is worth 20 points
Every candidate is classified by shape, and the shapes detect different failures. Being absent from a round-up is a different problem from being absent beside a named rival, and a prompt list written in one sitting is usually all one shape.
| Kind | What tracking it detects |
|---|---|
| Round-up | Whether you appear when someone asks for the best options in your market |
| Situational | Whether you are cited in answers to the problem your customers actually describe |
| Capability | Whether you surface on the specific things you do that others do not |
| Head-to-head | Whether you are in the consideration set beside a named rival |
A shape you track nothing in is an unmeasured question about your business, which is why coverage is a scoring signal rather than a note.
Duplicates are flagged before you spend
Every candidate is compared against the prompts already on your project, and a near-duplicate is marked as one. That matters more than it sounds: a tracked prompt is a recurring nightly cost on every assistant, so the cheapest question this feature can find is the one it stops you tracking twice.
From a candidate to a measurement
Three actions, and all three are reversible:
Track. Adds the question to your checked prompts and files it under a topic cluster in the same step, so the question you decided to measure is also the question your content plan has to answer. From that night it is checked like any other prompt, and it costs credits on every future check. See AI prompts for what happens next.
Reject. Removes it from the queue and keeps it, so RankX AI never offers it again.
Restore. Brings a rejected question back when the market moves.
Both track and reject work in bulk, and the queue has three views: new, tracked and rejected. Past runs are kept, so a search you ran last week reopens rather than being retyped.
What it costs
35 credits a run, whatever the run returns, up to eight seeds. Not per question and not per seed.
A run that persists no candidates is not charged. If every question came back filtered out or already tracked, the credits are not taken, even though the provider call was made and paid for. That call stays our cost.
A repeat of a recent search is free. RankX AI keeps the questions behind a seed for thirty days (three days for a seed that returned nothing), and a run whose every seed, market and option matches a cached one does no fresh provider work and opens no charge at all. Change a seed, the market or an option and the run is charged in full. Search trends works the same way inside its own seven-day window. (This paragraph said the opposite for a few hours on 8 September 2026; the application's own query bar states the rule, and the code was re-read before this was corrected.)
Every rate is published in the credit cost reference.
Who can use it
Every plan, the trial included. There is no research add-on and no per-seat fee.
In an agency client portal, AI prompt research sits behind the same permission as AI visibility, because it feeds the prompt surface that permission already covers. Its two siblings, keyword research and search trends, are ungoverned. See what clients can see.
Driving this from an assistant
research_prompts runs a search and list_prompt_candidates reads the queue
back. Both carry the demand caveat in their own output, so an assistant reading
the result cannot report the figure as a measurement either.
"Research prompts for our two main services in the UK. Show me the ten with the best coverage score that we do not already track. Do not track anything yet."
See the tool reference.
Where to go next
- The AI prompt research feature page, for what it costs, who it is for and how a candidate is scored, in one place.
- AI prompts, for what a tracked prompt does every night.
- Keyword research, for the same job over what people type into a search box.
- Search trends, for whether a subject is moving before you commit to it.
Last updated
Issue reference
Every check the RankX AI Website Audit runs, with its severity, its weight, what it means and which audit paths can produce a verdict for it.
Keyword research
Discovering keywords in RankX AI, what each figure means, how the cache makes repeats free, and how discovery feeds rank tracking and topic clusters.