Prompt research · The missing number

Why does prompt volume not exist, and what can you use instead?

A keyword arrives with a number that tells you what to target. A prompt does not, no dataset behind one exists, and the substitutes on sale are keyword volumes with an undisclosed mapping step bolted to the front.

Topic in Prompt research12 min readUpdated 30 Aug 20267 sources, 6 primary and 1 vendor-described

Back to the roadmap Next: Where prompts come from

On this page

The short answer

No engine publishes prompt frequency and no public dataset of it exists, because the unit you would have to count is a sentence rather than a string: frequently multi-turn, and often carrying session context that never appears in the question itself. Anything sold to you as a prompt volume is the estimated search volume of a keyword that some product decided your prompt means, with two undisclosed error terms stacked in front of it. The instability goes deeper than the input: a July 2026 critical survey of 45 studies reports that on the surfaces where temperature could be controlled, repeated runs at temperature zero change 9 to 28% of decisions.⁴

Key takeaways

No engine publishes prompt frequency, and what is sold in its place is a keyword estimate with a hidden mapping step in front. Ask for the denominator, the panel and the date; real data answers each in one sentence.
The demand signal that survives is grammatical rather than numerical: across 55,393 trending queries, AI Overviews fired on 64.7% of question-form queries against 9.5% of the rest.¹
Write the panel as questions in your buyers' words, then measure a proportion over it and publish that with an interval and a date. One run reads as more precise than it is.

The absence

Why is there no prompt volume number anywhere?

Because no engine publishes prompt frequency, and no one else is positioned to count it. Google describes exactly what it does publish: “sites appearing in AI features (such as AI Overviews and AI Mode) are included in the overall search traffic in Search Console”, reported inside the Performance report under the “Web” search type, aggregated with ordinary search and never itemised by the question that produced them.² Search Console’s own help then removes the part you most want: on the query tab, anonymized rare results are omitted from the table, and the table displays at most 1,000 rows, removing exactly the long tail where question-shaped demand lives.³ The other engines publish nothing comparable at all.

Four structural properties would defeat aggregation even if the logs were handed to you. Cardinality: a prompt is a sentence, and two buyers with identical intent routinely write sentences that share almost no vocabulary, so the distribution has an enormous tail of near-unique strings instead of a countable head. Session context: the sentence that triggered the answer may be “what about the cheaper one?”, whose meaning lives three turns earlier and never appears in the text. Personalisation: stored memory and account history change what the same words retrieve for different people. Non-determinism: the 2026 critical survey reports that on temperature-controllable surfaces, repeated runs at temperature zero change 9 to 28% of decisions, and that a four-engine study over 45 days saw daily source-level Jaccard scores of roughly 0.34 to 0.42.⁴ A frequency table needs a stable unit of observation, and a prompt is not one.

The proxy

What is a vendor's prompt volume actually made of?

A vendor’s prompt volume is a keyword search volume with a mapping step bolted to the front: the product decides which keyword a prompt “means”, then reports that keyword’s estimated volume as the prompt’s.

Error term one: the keyword estimate

Keyword volume is itself a model: a clickstream panel projected to a population, smoothed over a rolling window. It was never a count, and its confidence interval is not published by anyone who sells it.

Error term two: the mapping

Deciding that a nineteen-word conversational question “is” a three-word keyword is a judgement call made by a classifier. Its accuracy is measurable in principle and disclosed by nobody in practice.

The population that is missing entirely

The people you most want to count, those who asked an engine, got an answer and never searched afterwards, leave no trace in any keyword corpus by construction, so no mapping can recover them.

Three questions separate an honest proxy from a decoration, and any vendor can answer all three about a real dataset in one sentence each. What is the denominator: prompts on which engine, over which window? What is the panel: whose behaviour, how many people, recruited how? What is the date, given that these systems move month to month? A supplier who says “we estimate this prompt’s demand from the search volume of the keywords it maps to, measured in June 2026” is selling a proxy, which is a legitimate product. The failure mode is not the proxy; it is printing the word volume next to it and letting a founder plan a quarter around a number with an unbounded error term.

What it would take

What would a real prompt-volume dataset require?

One of three things, and each has an obstacle that is structural rather than temporary. An engine publishing its query logs: prompts are the most commercially sensitive and most privacy-loaded asset these companies hold, and none of them has published even an aggregate distribution. A consenting panel of real conversations: this is how keyword volume is estimated, and it transfers badly: a conversation is far more identifying than a click stream, and the cardinality problem means a panel would have to be orders of magnitude larger to stabilise any individual phrasing. An intermediary that sees the traffic: analytics and infrastructure providers see the click that follows an answer, if there is one, but never the sentence that produced it.

What does exist, and is worth knowing about, measures a different thing: how often an AI answer appears at all. A representative sample of 11,500 queries presented at SIGIR 2026 found AI Overviews shown for 51.5% of them.⁵ That is surface coverage on one engine, not demand, and the two are easy to confuse because both are percentages with the word “AI” nearby. Coverage tells you how often the format appears. Demand would tell you how many people asked. Only the first has ever been measured.

The one signal

Which demand signal survives all of this?

The question form survives, and it is the best-corroborated structural finding in this field. In a study of 55,393 trending queries collected over 40 days between 13 March and 21 April 2026, Google AI Overviews fired on 64.7% of queries phrased as questions and 9.5% of queries that were not, which the authors describe as a 6.8 times difference, against 13.7% across all queries.¹ The all-query baseline already contains the questions, which is why the non-question rate sits below it rather than equal to it. The paper’s own counts make the shape plain: of the 55,393 queries sampled, 4,198 were question-phrased, 7.6% of the sample, and 2,716 of those returned an AI Overview.

An independent method points the same way. A vendor study published in January 2026, drawing on 146 million search result pages and 1.9 million AI Overview citations, reported AI Overviews on 57.9% of question queries against 15.5% of non-question result pages, and on 46.4% of queries seven words or longer.⁶ The absolute numbers differ from the academic ones because the samples and query populations differ. The direction does not, and two dissimilar methods agreeing on direction is a stronger basis for a decision than either number alone.

Reading it correctly

How do you use the question signal without overreading it?

Read the question-form finding as a statement about when an AI answer appears on one engine’s results page, and never as a measure of how often anyone types a particular sentence.

What the finding is	What it is not	What follows for a panel
A trigger rate for AI Overviews on Google result pages	A trigger rate for assistants that always answer	Write panel entries as questions; on chat surfaces the analogous question is whether the answer retrieved the web at all
A property of the query’s grammatical form	A property of any individual question’s popularity	Use it to choose the shape of your entries, never to rank them
Measured on trending queries over 40 days in spring 2026	A constant	Re-check the direction annually; treat the level as stale within months
Evidence that AI answers appear more often on questions	Evidence that appearing is good for you	An answer firing is also the moment a click may not happen

The practical consequence is a change of research method rather than of tooling. Prompt research is qualitative research with a quantitative output, so the validity of the number you publish rests on the sampling frame rather than the sample size, the opposite of how keyword research works, and the reason a longer prompt list is not automatically a better one. The prompt research stage covers the panel and its arithmetic; the ranked list of places real prompts come from is in where prompts come from.

The substitute

What do you measure instead of a volume number?

You measure your own panel, the one denominator you control: fix the list of prompts, run it on a schedule across the engines that matter to you, record whether your domain was mentioned or cited, and report the result as a proportion of that panel. That is the opposite of a volume claim, because it describes a stated set of questions over a stated window rather than a market, and it stays honest exactly as long as the frame is published beside the number.

The arithmetic that matters here is repetition rather than size. Because the answers move between runs, a single pass reports a point where there is really a distribution. A March 2026 preprint sampled three generative search platforms repeatedly and found many apparent differences between domains falling inside the noise floor, concluding that visibility metrics belong in print as sample estimates with uncertainty rather than as fixed values.⁷ The critical survey turns that into an instruction: it cites a four-engine study proposing seven to eight repetitions per prompt as a starting point, warns that this is no universal standard because it comes from a small universe of Swiss queries, and recommends repeating the measurement until the interval around the estimate is narrow enough for the decision in front of you.⁴

Two habits follow. Re-run the same panel rather than assembling a fresh one each quarter, because changing the frame moves the number more than the market does. And put the interval beside any movement before calling it progress: on the forty-prompt, five-run panel that the prompt research stage works through, the band around each rate swallows several points of apparent change.

The honest limit of this article

This article argues from an absence, and an absence is hard to prove. The strong claim, that no engine publishes prompt frequency, is verifiable today and could stop being true the week after this page was updated. The 64.7% and 9.5% pair comes from a preprint rather than peer-reviewed work, was collected on trending queries rather than a random sample of all queries, and measures one engine’s results page. The vendor corroboration cannot be inspected: its sample is disclosed, its method is not, and the company that ran it sells tools whose value rises with the answer it found. The stability figures carry the survey’s own caveat: the Jaccard range comes from a small universe of Swiss queries. None of that changes the direction, and all of it should stop you quoting any of these numbers as a constant.

Where a product fits, and where it does not

Nothing in this article needs a purchase, and the substitute for the missing number is free: write your panel as questions in your buyers’ words, keep the non-question entries deliberate and few, and record the sampling frame beside the number you report. Bavior cannot tell you a prompt’s volume, does not estimate one, and does not sell a number of that shape, because the dataset that would justify it does not exist. What it does is later in the sequence: it runs a fixed prompt panel across five engines on a schedule and records which sources each answer cited, so you can see whether your position moved on the questions you chose; where a cited source is a live discussion thread, it drafts a reply on an account you control, which you approve before anything posts. The free AI visibility check and the free GEO audit run without a paid plan; paid plans are from $99/mo billed monthly, or $79.17/mo billed annually (as of 30 Aug 2026).

Sources, all checked 30 Aug 2026

Xu, Iqbal & Montgomery, 2026; 55,393 trending queries, 13 March to 21 April 2026; AI Overview activation 13.7% across all queries, 64.7% for question-form queries, 9.5% for non-question queries; Table 4: 4,198 question-phrased queries, 2,716 with an AI Overview (preprint): arxiv.org/abs/2605.14021
Google Search Central, “AI features and your website”, last updated 10 Dec 2025 (AI-feature traffic included in overall Search Console traffic and reported under the “Web” search type; first-party): developers.google.com/search/docs/appearance/ai-features
Google Search Console Help, “Performance report (Search results): Troubleshooting data discrepancies” (1,000-row table maximum; anonymized rare results omitted from the query tab; first-party): support.google.com/webmasters/answer/17010575
“Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026)”, 15 Jul 2026, arXiv:2607.14035 (survey preprint; 9–28% of decisions change across temperature-zero repeats; daily source-level Jaccard 0.34–0.42 from a four-engine, 45-day study on a small Swiss query universe; seven to eight repetitions per prompt suggested): arxiv.org/abs/2607.14035
Grossman et al., SIGIR 2026; representative sample of 11,500 queries; AI Overviews shown for 51.5%: arxiv.org/abs/2604.27790
Question-query study, January 2026: 146 million search result pages and 1.9 million AI Overview citations; AI Overviews on 57.9% of question queries against 15.5% of non-question result pages, and 46.4% of queries of seven words or more. Vendor-published; described rather than linked, per this curriculum’s sourcing rule.
Sielinski, “Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement”, Mar 2026, arXiv:2603.08924 (preprint; repeated sampling across three platforms, many apparent differences inside the noise floor, so a panel rate needs an interval): arxiv.org/abs/2603.08924

Continue reading

Learn GEO GEO roadmap Learn GEO Previous: Do GEO tactics work Learn GEO Next: Where prompts come from

FAQ

Frequently asked questions.

Can I get prompt volume from an AI engine's own analytics?

No engine exposes prompt-level frequency to site owners, and the closest thing published is aggregated beyond recognition. Google's documentation states that sites appearing in AI features are included in the overall search traffic in Search Console and reported under the "Web" search type, so AI-surface impressions arrive mixed with ordinary search and are never itemised by the question asked. Search Console additionally omits anonymized rare results from the query tab and caps the table at 1,000 rows, which removes precisely the long tail where question-shaped demand lives. The other assistants publish no site-owner reporting comparable to this at all.

If prompt volume does not exist, how do I decide which prompts to track?

Choose them by observed evidence that a buyer asked them, then size the panel statistically rather than by demand. The selection rule is attestation: every entry should trace to a recorded call, a support ticket, a query in your own search console, a community thread, or a stated objection, because those are the only places where a real person's question is written down with a date attached. The sizing rule is separate and arithmetic: presence in an answer is a binary draw, so 40 prompts run 5 times each per engine gives roughly a plus-or-minus 7-point band at panel level, which is what Prompt research works through.

Should every prompt in my panel be phrased as a question?

Most of them, and the exceptions should be deliberate rather than accidental. The evidence for the question form is unusually consistent: an academic study of 55,393 trending queries collected in spring 2026 measured AI Overview activation at 64.7% for question-phrased queries against 9.5% for non-question queries, and a separate vendor dataset of 146 million result pages measured 57.9% against 15.5% by a different method. Keep non-question entries where buyers genuinely type fragments, such as a bare category name or a product comparison written as two nouns, and label them, so a class that behaves differently is never averaged into the rest.

Does a high AI Overview trigger rate mean the question is worth targeting?

A high trigger rate means an AI answer is likely to appear, which is not the same as the question being valuable to you. The answer appearing is also the moment a click may not happen, so a question with a high activation rate can be worth targeting because you want to be named in the answer, or worth avoiding because you would rather have the visit. Decide per question, using the buying stage rather than the trigger rate: an objection a prospect asks before contacting you is worth being named in even if nobody clicks, while a comparison you currently win the click on may be worth leaving alone.

Bavior Editorial

The team that researches and maintains Bavior’s writing on Reddit marketing and AI search visibility. Every figure here is attributed to a named source with the date it was checked, and none of our links are affiliate links.

Found a number that looks wrong? Tell us and we will re-check it: support@bavior.com