Is AI Prompt Volume Data Estimated, Unreliable, or Made Up?
AI prompt volume data is estimated, not measured. The AI companies do not publish how often people ask a given question, so every prompt volume figure is built from a sample (paid panels, browser-extension data, or a vendor's own users) and then scaled up to the whole market. It is rarely made up, but it is often shown with more precision than the sample can support, so use it to compare topics, not to forecast traffic.
Why is there no "real" prompt volume?
Keyword volume works because Google measures searches on its own engine and shares aggregated counts through tools like Keyword Planner and Search Console. Nothing equivalent exists for AI assistants. OpenAI, Anthropic, Google's Gemini app and Perplexity do not publish how many people asked a particular question.
Even if they did, the unit would not behave like a keyword. Search queries are short and repeat exactly: millions of people type "best crm for small business". Prompts are long, conversational and personal: "we're a 12-person agency moving off spreadsheets, need something that handles retainers and integrates with QuickBooks, what should we look at?" Almost no two prompts are identical, so a "volume" for a prompt only exists after someone has grouped thousands of different wordings into one topic. The grouping is a modelling choice, and a different grouping gives a different number.
Where do prompt volume numbers come from?
Vendors rarely publish their full method, but the data can only come from a few kinds of source:
| Source | How it works | Main weaknesses |
|---|---|---|
| Paid or opt-in panels | People are paid to share chat history, answer surveys or write prompts | Small, and skewed toward tech-savvy people willing to be paid; mobile and workplace use often missing |
| Browser-extension or clickstream data | Usage data bought from data brokers, collected by extensions on users' desktops | Consent concerns; desktop-browser only; the supply can disappear after a browser or policy change |
| A vendor's own users | Conversations or searches from people using the vendor's product | Reflects that product's audience, not the whole market |
| Modelling from search data | Search keyword volumes converted into estimated AI demand with ratios | Assumes AI demand follows search demand, which is often untrue for complex questions |
Most products combine several of these and then apply a model to scale the sample to the full user base of each AI platform. That last step is where the uncertainty grows: a small sample, grouped into topics and multiplied up, can produce a confident-looking figure like "2.1 million on ChatGPT" that carries a wide unstated margin of error.
Vendors describe their data in different ways. Profound, for example, says its Prompt Volumes feature draws on more than 1.5 billion AI conversations licensed from several double-opt-in consumer panels, and can be filtered by platform, age range and income bracket; it does not publish panel sizes per country or a margin of error for a single figure. Conductor, which sells a competing platform, argues that no provider has direct access to the AI companies' data, so panel and extension-based estimates are statistically weak. Both are commercial positions, so judge any vendor by the answers to the questions below rather than by its marketing.
Is it estimated, unreliable, or fabricated?
Each word fits a different part of the problem:
- Estimated: yes, always. There is no source that counts every prompt, so every figure is an estimate.
- Unreliable: often, at the level shown. Absolute numbers, exact platform splits and demographic breakdowns stretch small samples the furthest. Relative comparisons between broad topics hold up better.
- Fabricated: rarely. Most vendors do have underlying data. The problem is usually presentation: an estimate displayed like a measurement, with no margin of error.
A good rule: the more specific the number (one prompt, one platform, one age band, one month), the less you should trust it.
Can you measure keyword search volume inside LLM prompts?
Not directly. There is no count of how often a word or phrase appears in prompts across ChatGPT or Claude, and exact phrases rarely repeat in long conversational prompts. What you can measure are proxies that point in the same direction:
- Search volume for the underlying topic. If a topic has little search demand, it rarely has large AI demand either, although complex, multi-part questions shift toward AI faster than simple lookups.
- Long queries in Search Console. Queries of eight or more words that reach your site are often written like prompts and show the wording people use. See how to use Search Console long-tail queries to find AI prompts.
- "People also ask" and related questions. They show which follow-up questions people ask around a topic.
- Your own conversations. Sales calls, support tickets and chat logs show the exact questions buyers ask before buying, which is the demand that matters to you.
Combining these gives a ranking of topics by likely demand, which is usually all a content or tracking plan needs. More detail in how to use keyword research to estimate AI prompt demand.
What should you ask a vendor before trusting its numbers?
- What are the data sources, and was the data collected with users' consent?
- How large is the sample for my category and country, not for the whole product?
- Which platforms are covered, and does it include mobile apps and workplace accounts?
- How are different prompt wordings grouped into one "prompt" or topic?
- How is the sample scaled up to a platform-wide figure?
- What is the margin of error, or at least the range, for a single number?
- How often is the data refreshed, and does it change when the method changes?
- Can I see example prompts behind a volume figure?
A vendor that cannot answer the first four is selling a number without a method. A vendor that shows example prompts and ranges, and is clear about what it cannot see, is giving you something you can use carefully.
How should you use prompt volume data safely?
- Compare, don't count. Use it to see that topic A likely has more demand than topic B, not to predict that a page will get a certain amount of traffic.
- Prefer broad topics to single prompts. Estimates for "payroll for small businesses" are steadier than for one specific phrasing.
- Weight by buyer value. A low-volume question asked just before purchase is often worth more than a high-volume informational one.
- Cross-check with your own data. If a "high-volume" topic never comes up with customers, trust your customers.
- Keep your tracking questions separate. The questions you track should come from real buyers, not from whatever a volume tool ranks highest; see how to choose the right prompts to track and whether tools monitor fixed prompts or generate them.
What do people ask about prompt volume?
Will AI companies ever release prompt volume data?
Not as market-wide volumes so far. Microsoft shows site owners the grounding queries behind Copilot citations of their own pages in Bing Webmaster Tools, and Google Merchant Center gives merchants product term insights for AI Mode, AI Overviews and Gemini, but no AI company publishes how often a given question is asked across all users. If one does, it will change how every vendor estimate should be judged.
Are prompt volumes useful for choosing content topics?
Yes, as one input among several. They are most useful for spotting topics people discuss with AI that barely show up in search data, and least useful as exact numbers.
Why do two tools show very different volumes for the same topic?
They use different samples, group prompts differently and scale them up with different models. Large disagreements between tools are a sign of how uncertain the estimates are.
Should I pay extra for prompt volume data?
Only if the vendor can explain its sources, sample size and ranges for your category. Without that, free proxies such as search volume, long Search Console queries and your own sales conversations give you a similar ranking of topics.
Published by The Kovmere Team · updated Sep 28, 2026 · More resources