Prompt Volume: What It Is and How It's Measured

Every SEO tool built in the last eighteen months has quietly added a new number to its keyword screens. It sits next to monthly search volume, it is usually smaller, and it claims to tell you how often people ask an AI assistant about a topic. Some vendors call it prompt volume. Some call it AI search volume. DataForSEO calls it ai_search_volume. Whatever the label, it is being sold as the successor to the metric the entire SEO industry has planned around for twenty years.

It is a genuinely useful number. It is also, in every product that sells it, a model of something else. Nobody outside OpenAI, Google, and Anthropic can see what people actually type into an assistant, so no vendor is counting prompts. They are inferring them, from three completely different proxies, which is why two tools will hand you two different numbers for the same phrase and both will be defensible.

We sell one of these numbers, so this is an awkward post to write. ContextBolt SEO has an ai_keyword_volume lookup and it costs a credit like everything else. This piece explains where that number comes from, where the competing numbers come from, and how to use any of them without building a content plan on sand.

What is prompt volume?

Prompt volume is the estimated number of times people ask AI assistants about a given keyword or topic in a defined period, usually a month. It is the AI-search equivalent of monthly search volume, and it exists because the queries people bring to an assistant are not the queries they bring to a search box.

That difference is the whole reason the metric was invented. Nobody types “best crm software” into ChatGPT. They type four sentences about their team size, their budget, and the tool they are leaving. Keyword volume was built for short, repeated, near-identical strings. Prompts are long, personal, and almost never repeat word for word, so counting exact matches would return zero for nearly everything.

So prompt volume is not a count of a string. It is a count of a topic, reconstructed by clustering similar prompts and attributing them to a keyword. That reconstruction step is where all the disagreement lives.

Why keyword volume alone stopped being enough

The scale is the argument. In July 2025, OpenAI told Axios that ChatGPT users were sending 2.5 billion prompts a day, around 330 million of them from the US. Eight months earlier the figure was one billion. That is a channel with real demand in it, and until recently there was no keyword research surface pointed at it at all.

The second argument is what happens on the Google side. We measured this on our own property and published the numbers. Seven pages pulled 104,008 impressions at an average position of 6.6 and returned 395 clicks, a click-through rate of 0.38 percent, while worse-ranking pages on the same site returned 4.7 times more clicks per impression. The full breakdown is in AI overview click loss. Ranking is no longer the same thing as traffic, so a metric that only describes ranking demand describes less of the picture than it used to.

None of that makes prompt volume trustworthy. It just makes it worth understanding.

Nobody can see the prompt logs

This is the part most vendor pages skip past in a subordinate clause.

Google publishes query data through Search Console. It is first-party, it comes from the platform itself, and Google historically held over 90 percent of search, so a single reliable source covered nearly the whole market. That is the foundation monthly search volume was built on.

There is no equivalent for assistants. OpenAI does not publish a prompt report. Neither does Anthropic, Perplexity, or Google for its AI Mode conversations. Demand is also fragmented across at least five major assistants rather than concentrated in one. So every prompt volume figure on the market is an inference built from data the vendor could legally obtain, and the vendors could not obtain the same data.

That is not a scandal. It is just the constraint. It does mean that comparing two vendors’ numbers for the same keyword is a category error, because they are not measuring the same thing.

The three ways prompt volume is actually measured

Every product selling this metric uses one of three methods. Knowing which one you are buying tells you what the number is good for.

Method 1: consumer panels

Panel vendors recruit people who opt in to having their AI usage recorded, capture the real prompts those people send, cluster them, and extrapolate to the wider population. Profound is the clearest example. Its Prompt Volumes documentation describes tens of millions of real prompts a month from double opt-in consumer panels, with probabilistic modeling applied to correct demographic and geographic bias, covering 10-plus regions from January 2025 onward.

The strength: these are real prompts. Actual sentences that actual people sent to an assistant, which no other method can claim.

The weakness: the sample. Conductor’s teardown of the metric puts panel and extension coverage at far less than 1 percent of real traffic, possibly as low as 0.15 percent, and points out that opt-in panels skew technical and paid panels skew toward people motivated by the payment. Extrapolating from 0.15 percent to a national figure is a big multiplier on a biased base.

Method 2: People Also Ask modeling

The second approach never touches an assistant. It infers conversational demand from Google data that already exists in question form.

This is what sits underneath our own tool, so here is the mechanism in full. DataForSEO’s ai_search_volume is, in its own words, “the estimated frequency with which a specific keyword is used in questions that people may ask AI tools,” and its published methodology states that the figure comes from a proprietary algorithm whose signals include the People Also Ask section of Google results, drawn from their SERP index. The live endpoint returns the current month plus a 12-month trend, and accepts up to 1,000 keywords per request.

The strength: stability and coverage. PAA is a huge, consistent, long-running dataset, so the numbers do not swing wildly month to month, and you can pull thousands of keywords cheaply.

The weakness: it is a proxy for a proxy. It measures question-shaped demand as expressed on Google, then treats that as a stand-in for demand as expressed to an assistant. When those two diverge, and they will, the model will not notice. There are also documented quirks worth knowing. Grammatical variants score identically, so “tie” and “ties” return the same value, and multi-word phrases only count questions containing every word.

Method 3: your own first-party data

The unglamorous one. Your analytics already records referrals from chatgpt.com, perplexity.ai, and Google’s AI surfaces. Your server logs record which AI crawlers fetched which pages, which is worth cross-referencing against the AI crawlers list.

The strength: it is counted, not modeled. Zero extrapolation.

The weakness: it only shows demand that already reached you, so it is useless for finding topics you do not cover yet, and the absolute numbers are small enough to be noisy.

Prompt volume methods compared

What prompt volume is genuinely good for

Strip out the marketing and there are three jobs this metric does well.

Ranking topics against each other: if topic A scores 4x topic B in the same tool, that gap is probably real even if neither absolute number is. Relative order survives a systematic modeling error. Absolute magnitude does not.

Spotting phrasing you would never have guessed: panel data in particular surfaces the way people actually ask, which is longer and more specific than any keyword tool would suggest. That reshapes your headings more usefully than it reshapes your topic list.

Catching a trend early: a 12-month trend line on a modeled number is still a trend line. If a topic is climbing steadily across two independent tools built on different methods, something real is happening underneath both.

What it is not good for

Traffic forecasting. You cannot multiply prompt volume by a click-through rate, because assistants often answer without linking to anyone. There is no stable CTR curve for AI answers the way there is for the Google top 10.

Reporting a number to a client or a board. If you publish “12,000 monthly prompts” and a competitor’s tool says 800, you have a credibility problem you did not need. Report direction, share of voice, and citations instead.

Replacing keyword research. Google search still exists and is still measured far better. Prompt volume is a second lens, not a swap. Anyone telling you to retire monthly search volume in 2026 is selling something.

How to use prompt volume without getting fooled

Five rules that have survived contact with our own data.

1. Read the rank, not the value. Sort your keyword list by prompt volume and use the ordering. Ignore the digits.
2. Never mix two vendors’ numbers in one table. Different methods, different scales, no valid comparison. Pick one source per analysis.
3. Pair it with the Google number, always. A term with high search volume and near-zero prompt volume is a classic-SEO play. The reverse is an AEO play. That contrast is more useful than either figure alone.
4. Check the trend before the level. A 12-month direction is more trustworthy than a single month’s magnitude in every one of these methods.
5. Validate against your own logs. First-party AI referral data is small but it is real, and it is the only thing that tells you whether the modeled demand ever showed up.

Should this change what you publish?

Mostly it should change your headings, not your topic list.

The consistent finding across every method is that prompts are question-shaped and longer than queries. That is an argument for the structural work covered in answer engine optimization: question-form H2s, self-contained paragraphs that survive being pulled out of context, checkable specifics an assistant can quote, and a clean definition of the term you want to own. Chunk retrieval decides what gets cited, and a page organized around exact questions gets chunked more usefully than a page organized around a marketing narrative.

If you want to see where you currently stand before optimizing anything, tracking citations is the more actionable starting point. We compared the field of tools that do it in best AI visibility tools.

The honest take

Prompt volume is a real metric describing real demand, sold with a precision nobody has earned.

The demand is not in question. Two and a half billion prompts a day is not a rounding error, and content built for how people ask assistants will beat content built for how people query Google in that channel. But every figure in every dashboard, including the one our own product returns, is a reconstruction. Panels reconstruct from a sliver of real prompts. PAA models reconstruct from Google’s question data. Neither has ever seen an assistant’s logs, and neither will.

So the useful posture is to treat prompt volume the way a sensible person treats a keyword difficulty score. It is a sorting aid. It is not a measurement. Use it to decide what to write next week, never to promise what that will return.

ContextBolt SEO returns the AI volume next to the Google volume for the same keyword, inside whatever agent you already work in, for a flat $35 a month with 1,000 lookups included. The ai_keyword_volume lookup costs 1 credit. We tell you which method sits behind it, which is more than most of this market does.