TL;DR: No AI provider publishes what people ask it, so nobody can tell you how often a question is asked. Every prompt volume figure on the market is modelled, usually by generating plausible prompts and running them through the models. That measures what the models answer, not what people ask, and it is useful as long as nobody mistakes it for demand data.
Key Takeaways
There is no keyword planner for assistants. Providers do not publish query data and show no sign of starting.
Vendor figures are estimates of an unobservable quantity, not measurements.
Synthetic prompts measure model behaviour, which is genuinely useful and is a different thing.
Your own sources of demand are better: sales calls, support, search data, community language.
A prompt set built from real conversations beats any volume estimate, because it reflects your buyers.
Ask one question of any vendor: where does the number come from, and what exactly was counted.
Anchor targets assume your CMS generates slug IDs from heading text. Confirm on staging and adjust the hrefs if your platform slugifies differently.
Why nobody has this data
Search volume exists because search engines chose to publish it, through advertising tools built to sell keywords.
The structural difference
Search | Assistants |
|---|---|
Query data published through ad tools | Not published |
Third-party panels approximate it | No equivalent panel of conversations |
Queries are short and comparable | Conversations are long and varied |
One query, one intent | A conversation shifts intent across turns |
Public results pages to scrape | Private, personalised sessions |
Why providers are unlikely to publish
The conversations are private, frequently personal, and the commercial incentive that produced search volume data, selling keyword advertising, does not currently exist in the same form.
That could change. It has not yet.
The counting problem
Even with access, what would you count? A conversation about project management software that mentions three products across eleven turns is not one query.
Search volume works because a query is a discrete unit. A conversation is not, which means any prompt volume figure has had to invent a unit of counting as well as estimate the number.
What that implies
Any figure presented as the number of times people ask something has two layers of assumption underneath: how it was counted and how it was estimated.
Neither is usually disclosed.
The honest position
This data does not exist publicly. Anything you are shown is a model's output, and the useful question is what the model was built from.
What this page is not arguing
That the vendors are dishonest or the tools are useless. Several do careful work. The argument is about what the number is, which the next sections cover.
What vendors are selling
Three approaches, with different amounts of reality behind them.
1. Synthetic prompt generation
Generate many plausible prompts for a topic, run them through the models, record what comes back. The brief describes this honestly in one paragraph and describes it as gathered user inputs in another.
What it measures: how models respond across a range of phrasings. Real and useful.
What it does not measure: how often anyone asks.
2. Search data as a proxy
Take search volume for related terms and model conversational demand from it, on the assumption that the two correlate.
The problem: the premise of this entire field is that people ask assistants differently from how they search. If the two correlate closely enough to substitute, that premise is weaker than the field claims.
3. Panels and opt-in data
Some vendors recruit users who share their conversations. This is the closest thing to real data and it is small, self-selected, and not representative in the way search panels are.
Worth more than the other two, and worth asking about sample size.
How to tell which you are looking at
Signal | Likely method |
|---|---|
Covers every topic instantly | Synthetic or search-derived |
Figures look like search volume | Search-derived |
Described as modelled or estimated | Honest synthetic |
Discusses sample size and recruitment | Panel |
Method not described at all | Ask |
The one that should concern you
Any tool presenting a clean monthly number with no stated method. That number has the shape of search volume and none of its basis, and it will be read as demand by everyone who sees it in a report.
The fair summary
These tools can tell you useful things about how models behave. They cannot tell you how many people ask, and the better vendors say so.
What synthetic prompts measure
This is the dominant method, and it produces something genuinely valuable once you stop calling it volume.
What it tells you
Finding | Value |
|---|---|
Which brands appear across many phrasings | High. This is the useful output |
How answers change with constraints | High. Budget, size, integration |
Which sources get cited | High. Your outreach targets |
How consistent the answers are | High. This is your variance |
How often humans ask any of it | None |
Four of those five are worth paying for. The fifth is the one on the marketing page.
The right name for it
Model behaviour testing. You are probing a system with inputs to see how it responds, which is a legitimate and informative exercise.
Calling it volume imports an assumption about human behaviour that the method cannot support.
Where the synthetic approach is weakest
The prompts are generated from what someone thinks people ask. If that assumption is wrong, every downstream figure inherits the error, and nothing in the method would reveal it.
That is why prompts drawn from real sales conversations are better inputs than generated ones, whatever the volume question.
The circularity worth noticing
If a vendor generates prompts using a language model, tests them against language models, and reports the results, the whole loop runs on model output. Human behaviour never enters it.
That is not useless, and it is a closed system.
How to use it well
Treat the output as a map of how assistants currently behave in your category. Act on the brands, the framings, the constraints and the cited sources.
Ignore the volume figures, or at least never put them in a document where someone will read them as demand.
Questions to ask a vendor
Five, in writing, before buying anything that reports a volume figure.
The questions
Where does this number come from? Synthetic prompts, search data, a panel, or something else. A clear answer is itself a good sign.
What counts as one prompt? A conversation, a turn, a session, a unique phrasing. There is no standard, so theirs is a choice they made.
If it is a panel, how many people and recruited how? Small and self-selected is still informative, and you need to know which.
How would I know if the number were wrong? The best question in this category. If there is no way to check, that is the answer.
What does this tool not tell me? The honesty question, which works across this whole field.
Reading the answers
Answer | Reading |
|---|---|
We generate prompts and test them, so it is an estimate of model behaviour | Honest. Buy it for that |
We model from search data | Ask why conversational demand differs then |
Proprietary methodology | Not an answer |
It reflects real user queries | Ask where those came from |
The one to listen hardest to
The fourth. A vendor who can describe how their estimate could be wrong has thought about it. One who says it is accurate has not, or is not telling you.
Why this matters internally
The number will end up in a slide. Someone senior will read it as demand, plan against it, and later ask why the forecast did not hold.
Knowing what it is now is considerably cheaper than explaining it then.
What a good vendor says
That it is modelled, what the model is built from, what it is good for, and what it cannot tell you. Several say exactly this, and they are the ones worth buying from.
What to use instead
You already have better evidence of what your buyers ask, and it is specific to them rather than modelled across a market.
The sources
Source | What it gives |
|---|---|
Sales discovery calls | The questions buyers ask, in their words, before deciding |
Support tickets | What confuses people after buying |
Your site search | What visitors looked for and did not find |
Search Console queries | Real queries, real counts, from a search engine |
Community threads | How people describe the problem to each other |
Your enquiry form | How they found you, if you added the field |
Why this beats a volume estimate
A market-wide figure tells you what an average buyer might ask. Fifteen discovery calls tell you what your buyers do ask, including the constraints they attach and the phrasing they use.
The second is more specific, more current, and free.
Search Console deserves emphasis
It reports actual queries with actual counts, from a search engine, about your site. In a field full of modelled numbers, that is first-party data and it is underused.
The queries people search are not identical to what they ask assistants, and they are closer than a generated prompt is.
The practical exercise
Listen to fifteen discovery calls. Write down every question asked in the first five minutes and every constraint stated, verbatim. That is an afternoon.
You will end up with twenty to thirty real questions, which is a better input to both content and measurement than any volume table.
What you lose
The ability to say a topic gets some number of prompts a month. You could not have said that honestly anyway.
What you gain
A list you can defend, that nobody can dispute the source of, and that reflects your actual market rather than a model of a generic one.
Building a prompt set
The practical replacement for volume data, and the foundation of any measurement you do.
Where the prompts come from
Real conversations. Fifteen discovery calls, read or listened to, with every question and constraint written down verbatim.
Not generated, not guessed, not taken from a keyword tool.
The five types to include
Type | Example shape |
|---|---|
Category | What tools do X |
Comparative | X versus Y for a given purpose |
Constrained | X for a team of this size, with this budget, that integrates with Z |
Problem-first | How do I stop this thing happening |
Branded | What is your company, is it any good |
The constrained ones matter most. They are how buyers actually ask, they are where smaller companies can appear, and they are the hardest for a generated prompt set to produce convincingly.
How many
Twenty is enough to start and small enough to actually run. Growing it to fifty over time is reasonable; starting at fifty means nobody runs it twice.
Keep them fixed
The set is the instrument. Changing prompts between runs makes the comparison meaningless, so additions go in a separate group and the core set stays stable.
Review it quarterly
Add prompts from new sales calls, retire ones nobody asks any more. Note the changes with dates, so a later reader knows the instrument changed.
What this gives you
Appearance rates across your own questions, with ranges, that you can act on. Not market demand, which nobody has, but something better for deciding what to write and where you are losing.
The comparison worth making
If you also have a vendor's volume data, compare their top topics against your discovery call list. Where they disagree, trust the calls.
Will this metric ever exist
Worth thinking about, because the answer shapes how much to invest in the alternatives.
What would have to change
Condition | Likelihood |
|---|---|
A provider publishes aggregate query data | Possible, if advertising arrives |
An advertising product creates the incentive | The most plausible route |
A standard unit of counting emerges | Needed either way |
Panels reach representative scale | Hard, given how personal conversations are |
The advertising route
Search volume data exists because search engines sell keyword advertising and advertisers need to know what to bid on. If assistants develop a comparable advertising product, something like query data may follow.
That is the clearest path and it is speculative.
The counting problem persists
Even with publication, somebody has to decide what one unit is. A conversation that wanders across three products and eleven turns does not decompose into queries cleanly, so any published figure will carry a definitional choice.
When it arrives, the first question will be the same as the one this page recommends: what exactly was counted.
What to do meanwhile
Build the thing that will still be useful either way: a prompt set from real conversations, run consistently, with your own baseline and variance.
If volume data arrives, that set becomes more valuable rather than less, because you will have a history to compare against.
The thing not to do
Wait. Teams delaying measurement until proper data exists will have no baseline when it does, and the comparison they most want will be unavailable.
The honest forecast
Nobody knows. The page's position is that you should not plan around it arriving, and you should not be surprised if it does.
Six mistakes
1. Treating a prompt volume figure as demand. No provider publishes query data, so every number on the market is modelled. It will still be read as demand by whoever sees the slide.
2. Not asking what counts as one prompt. There is no standard unit. A conversation spanning eleven turns and three products is not one query, and the vendor chose how to count it.
3. Generating your prompt set. Prompts built from assumptions inherit those assumptions, and nothing downstream reveals the error. Use real sales calls.
4. Dismissing the tools entirely. Synthetic testing genuinely shows which brands appear, how answers shift with constraints, and which sources get cited. Buy it for that.
5. Ignoring Search Console. It reports real queries with real counts about your own site. In a field of modelled numbers, that is first-party data.
6. Waiting for proper data. If it arrives you will have no baseline to compare against, which is the comparison you will most want.
The pattern
Most come from wanting a figure shaped like search volume badly enough to accept one that has its form and not its basis.
The one to do this week
Listen to fifteen discovery calls and write down every question asked in the first five minutes, verbatim.
An afternoon, and it produces a better prompt set than any tool will sell you, built from the only demand data that is definitely real.
FAQs
Is there a keyword planner for AI assistants?
No. Providers do not publish query data, and the advertising incentive that produced search volume figures does not currently exist in the same form.
Where do prompt volume numbers come from then?
Usually synthetic prompts generated and tested against the models, sometimes modelled from search data, occasionally from small opt-in panels. All are estimates of something unobservable.
Are those tools useless?
No. Synthetic testing shows which brands appear across phrasings, how answers change with constraints, and which sources get cited. Those are worth paying for. The volume figure is the weakest output.
What should I ask a vendor?
Where the number comes from, what counts as one prompt, the panel size if there is one, how you would know if it were wrong, and what the tool cannot tell you.
Why is counting so hard?
Search works because a query is a discrete unit. A conversation spanning eleven turns and three products is not, so any figure carries a definitional choice as well as an estimate.
What should I use instead?
Fifteen discovery calls, support tickets, site search, Search Console queries and community language. All specific to your buyers rather than modelled across a market.
How do I build a prompt set?
From real conversations, verbatim. Twenty prompts across five types, weighted toward constrained questions, kept fixed between runs so comparisons mean something.
Will this data ever exist?
Possibly, most plausibly if an advertising product creates the incentive. Build your own measurement now regardless, because without a baseline you will have nothing to compare against.
Listen to fifteen calls
Take fifteen recent discovery calls and write down every question asked in the first five minutes, plus every constraint stated, in the buyer's own words.
An afternoon. It produces a better prompt set than any tool will sell you, and it is built from the only demand data that is definitely real.
Then
Sort them into five types: category, comparative, constrained, problem-first, branded.
Keep twenty, weighted toward the constrained ones, since that is how buyers actually ask.
Run them across three assistants, three times each, and record ranges.
Record the cited sources as well as who appears. That list is your outreach targets.
Keep the set fixed between runs, with additions in a separate group.
Review quarterly from new calls, noting changes with dates.
If the gap turns out to be content
It usually does, once the measurement names the questions you lose. Blazly SEO covers research, planning, drafting and direct publishing from about $39 a month.
Summary
No AI provider publishes what people ask, so there is no keyword planner for assistants and every prompt volume figure on the market is modelled. Most come from generating plausible prompts and running them through the models, which measures how the models respond rather than how often anyone asks, and which is genuinely useful under its real name. Even with access, somebody would have to decide what one prompt is, since a conversation spanning several turns and several products does not decompose into queries cleanly. Ask any vendor where the number comes from, what was counted, and how you would know if it were wrong. Meanwhile your own sources are better and free: discovery calls, support tickets, site search and Search Console, which reports real queries with real counts.