There Is No Keyword Volume for AI Prompts — Here's What to Use Instead

Search marketing has been built for twenty years on a single number: how many people search this term each month. It came, ultimately, from Google, which had the data and chose to expose a version of it. Every keyword tool, content plan and traffic forecast in the discipline rests on that foundation.

That number does not exist for AI prompts, and it is not coming.

This is not a temporary gap waiting on better tooling. It is structural, and understanding why is the difference between building a prompt strategy that works and buying a dataset that is quietly fictional.

Why the number cannot exist

The data is private and there is no commercial reason to release it. Google published keyword volume because it sold ads against those keywords — advertisers needed volume estimates to bid. AI assistants overwhelmingly do not sell keyword-targeted ads. There is no advertiser to serve, so there is no product reason to expose the data, and considerable privacy reason not to. Conversations with an assistant are markedly more personal than search queries, and a substantial share of what people ask would be identifying if released at any granularity.

There is no shared surface to scrape. Search rankings are public: anyone can query Google and observe the SERP. That observability is what made third-party volume estimation possible in the first place — you could scrape results, model click curves, and calibrate against known data. An AI conversation is private between the user and the assistant. There is no public artifact to sample.

Prompts are not a finite set. Keywords cluster because people type short, similar strings. Prompts are conversational, long, personal, and frequently multi-turn — the user refines across three messages, and the “prompt” that mattered was the conversation, not any single string. There is no canonical form to count.

And it fragments across seven-plus engines, each with a different user base and different behaviour, with no cross-vendor aggregator.

Put together: no willing publisher, no scrapable surface, no stable unit, no aggregation. Every one of those would have to change.

What the numbers you have been shown actually are

Vendors do show prompt volume figures. They are produced one of three ways, and the honest version of each is worth knowing.

Search volume with a question stapled on. Take the Google volume for “best crm,” relabel it “how do I choose the best CRM,” present the search number. This is the most common method. It is not a prompt volume; it is a search volume wearing a costume, and it inherits every way search demand differs from assistant demand.

Panel extrapolation. Recruit users, observe their assistant usage, extrapolate. Methodologically legitimate and severely limited by sample size and self-selection. People who agree to have their AI conversations monitored are not a random sample of people who use AI, and the topics they are willing to be monitored on are not a random sample of topics.

Model-generated estimates. Ask a language model how often people ask something. This produces a fluent, confident, entirely unfounded number. The model has no access to query logs. It is generating a plausible figure, which is exactly what it does with every other number it does not know.

None of these is fraud. All three become misleading the moment the output is presented as measurement rather than estimate — and by the time a figure reaches a slide it has usually lost its error bars.

The question to ask any vendor showing you prompt volume: where does the underlying observation come from? There are only three possible honest answers — a panel, a public proxy, or a model — and each implies a specific limitation. A vendor who cannot answer crisply is showing you the third one. This is one item on a longer list in how to audit an AI visibility vendor’s numbers.

What to use instead

The absence of volume data is less crippling than it first appears, because volume was never the thing you actually wanted. You wanted to know which questions precede a purchase. Volume was a proxy — a good one, because it was cheap and measurable. Several other proxies are available, and some are better.

Your own Search Console queries. The highest-value substitute. These are measured, first-party, and in your customers’ own words — with the caveat that Google withholds rare queries entirely, as its performance data documentation explains, so what you see is the head of your demand rather than all of it. They are terse where prompts are conversational, but intent maps cleanly: “crm for startups” and “what’s the best CRM for a startup” are the same question in two registers. This is a real dataset about real demand from your real audience, which is more than any prompt-volume vendor can offer. Search Console is the most underrated AEO dataset covers reading it properly.

Your sales and support conversations. The single most underused source in AEO. Your sales team hears the actual buying questions, phrased conversationally, every day — that is exactly the register a prompt is in. Support tickets give you the post-purchase half. Neither has a volume figure and both have something better: confirmed commercial relevance.

People Also Ask and autocomplete. Google’s People Also Ask boxes are question-shaped by construction and derived from real query data. Not volume, but a strong signal that a question is asked enough to be worth surfacing.

Community threads. Reddit, forums, Q&A sites, industry Slacks. People ask assistants the same things they ask communities, in the same conversational register. Engagement — upvotes, replies, recurrence — is a rough demand proxy.

Competitor visibility. Backwards but effective: if a competitor is being named in AI answers for a set of questions, those questions are being asked and are commercially live. Competitor benchmarking makes this observable rather than anecdotal.

Building a prompt set without volume

Since you cannot rank prompts by demand, rank them by something you can actually defend.

Commercial proximity first. How close is this question to a purchase decision? “Best CRM for a 10-person sales team” is adjacent to a buying decision. “What is a CRM” is not. In the absence of volume, proximity to revenue is the better sort key anyway — arguably better than volume ever was, since volume systematically favours the top of the funnel where it is cheapest to be irrelevant.

Cover the decision, not the topic. A buyer asks a sequence: what are my options, how do they compare, is this one right for my situation, what do users say, what does it cost. Track across that arc rather than piling up variants of one question.

Include the questions you are afraid of. “Is [your brand] worth it,” “[your brand] alternatives,” “problems with [your brand].” People ask assistants these constantly and the answers are frequently unflattering and unmonitored. Nobody enjoys adding them; they are consistently the highest-value prompts in a tracked set.

Keep it small enough to read. Twenty to fifty prompts you actually review beats five hundred you scan as a number. Each scanned prompt costs credits across every engine, and an unread result is a purchased number with no reader. Choosing tracked prompts works through the sizing trade-off.

Then let the data reorder itself. Once scanning, you learn which prompts are volatile, which are competitive, which are locked up by an incumbent. That is real evidence about your own market, generated by your own monitoring, and after a quarter it is a better prioritisation input than any estimate you could have bought.

The counter-argument

The fair objection: without volume you cannot forecast, and without a forecast you cannot build a business case, and without a business case AEO does not get funded.

That is true and it is a genuine cost. The honest response is not to invent the number — it is to build the case on a different footing. The strongest available framing is share-based rather than volume-based: of the questions we know matter to our buyers, we appear in this fraction, and our competitors appear in this larger one. That is measurable, defensible, and moves in response to your work. Building the AEO business case develops it, and measuring GEO/AEO ROI covers what can and cannot be attributed.

A relative metric you can actually measure beats an absolute metric you had to fabricate — and it has the additional advantage of not collapsing the first time someone asks where the number came from.

The summary

There is no keyword volume for AI prompts. Anyone selling you one is selling a proxy, a panel, or a guess, and the professional move is to ask which.

Meanwhile the inputs you need are sitting in Search Console, in your CRM notes, and in your support queue — measured, first-party, and specific to your actual buyers. That is a better starting position than the discipline had in 2005, when everyone was reading the same public volume figure and competing on who could act on it fastest.