AI search volume, tested: what the number actually counts

Every AI visibility tool now shows a number for how often people ask AI assistants about something. We pulled one vendor's AI search volume for 308 phrases and set it against Google's own. It is not a count of AI conversations: it ranks phrases almost exactly as Google does, counts words rather than phrases, and inflates any phrase that is also an ordinary word. Used for what it is, it is still useful.

What we found

• It is modelled, not measured. The vendor we tested builds the figure from the questions Google shows in People also ask. No vendor we read publishes a count of real AI conversations for a phrase, and two of the largest say plainly that no reliable source exists.
• It ranks like Google. Over 100 head terms and brands, its order agreed with Google's at 0.92, and its size sat at 3% of Google's at the median.
• It counts words. A phrase's figure is every question containing all of its words, in any order: reversing the words changed nothing in 60 of 60 cases, and adding a word never raised a figure (0 of 68 steps).
• So full questions read close to nothing. Of the 30 questions eight words or longer, 3 had any figure at all.
• Ordinary words inflate it. Brands that are also words — Wave, Monday, Shark — read ten to twenty times the usual ratio to Google, because every question containing the word counts.
• Its months are not a trend. Across the whole set the figure fell by a third between autumn and summer in both markets at once, while Google's was flat.

Where the number comes from

We tested DataForSEO's AI Keyword Data, the figure Recomma itself shows. Its documentation says the figure is "calculated using statistical data from questions in the 'People Also Ask' SERP element", reflecting "estimated usage in AI tools", and its help centre that a multi-word phrase counts the questions that "contain all the specified words". So the figure starts from Google, which is worth knowing before reading it against Google.

The others take different routes, and are candid about the gap:

What is public is the scale. OpenAI told Axios in July 2025 that ChatGPT handled 2.5 billion prompts a day, and OpenAI's own study of how people use it put seeking information at about a quarter of conversations. Nobody publishes those prompts broken down by phrase, which is the number every tool would need.

It ranks like Google

For the 100 head terms and brands with a figure on both sides, the AI figure ran at 3% of Google's monthly searches at the median, with half of them between 2% and 10%: 6% for head terms, 2% for brands. Its order agreed with Google's at 0.92 on a scale where 1 is the same list. The United Kingdom came out at 18% of the United States, against 20% for Google, over 103 phrases.

That agreement is the useful part and the limit at once. A figure that tracks Google this closely will tell you which of your prompts sits on more demand. It cannot tell you that people ask an assistant about something Google never sees, because it starts from what Google sees.

It counts words, not questions

We built ladders from a head term to a long tail, one or two words a rung. If the figure counted a phrase, a longer rung could sometimes read higher; if it counts every question containing the words, it never can. It never did — 0 of 68 steps — where Google's own figures rose on 5 of 57.

Word order is ignored altogether. "best crm software" and "software crm best" both read 928. All 60 reversed phrases matched their originals exactly; Google gave the same figure for 7 of the 38 it had either way round.

The consequence for anyone tracking prompts: the questions a tracker stores are sentences, and sentences match almost nothing. We asked for the 100 buying questions from our interface-versus-API study verbatim. 43 of the 70 under eight words had a figure; 3 of the 30 longer ones did. A prompt has to be read by a short phrase behind it, or it reads zero.

Ordinary words break it

Counting every question that contains a word goes wrong when the word means more than one thing. These phrases ran at more than ten times the median ratio to Google:

Wave the accounting software is caught up in every question about waves; Shark the vacuum in every question about sharks. Dreame reads hundreds of times its Google volume — consistent with the vendor folding word forms together, so that it counts questions about dreams. The robot-vacuum-mop and hosting phrases are a gentler case of the same thing: common words that appear together in questions about something else.

The months move on their own

Each figure comes with twelve months of history, and it is tempting to read it as a trend. Summed across the set, with the ambiguous phrases left out, the AI figure fell to 66% of its October 2025 level by the summer and recovered in September — in the United States and the United Kingdom together — while Google's moved within its usual range:

Phrase by phrase, the AI figure's months correlated with Google's at a median of 0.27, and 72 of 77 phrases read lower in the last three months than the first three, against 40 on Google. A fall that size, shared by running shoes and password managers at once, is the source changing rather than people asking less.

How we use it in Recomma

Recomma shows the figure on the Prompts page as Questions/mo, and the study changed how:

• A short phrase, never the sentence. Each prompt is read by the keyword it was drafted from, or the question Google showed people asking, or — for a prompt typed by hand — its topic. A topic's figure is greyed and says how many prompts share it.
• Named for what it is. "Questions/mo", with a ~ on every figure and Google's volume beside it, rather than a name that claims a count of AI conversations.
• Broad phrases are flagged, not ranked. Where the figure runs above 30% of Google's for the same phrase — ten times the median here — the column says "broad" and the row sorts with the unread ones.
• No trend line. The months are kept, but not drawn under each figure as if they were demand.
• Read for comparison. Which of your prompts sits on more demand is the question it answers well. How many people ask an assistant this exact question is one it cannot.

The column is documented in the Prompts guide.

Method and limits

• Phrases. 308 in the ten categories of our interface-versus-API study: its 100 questions verbatim; three ladders per category from a head term to a long tail; the two- and three-word rungs with their words reversed; and five brands per category.
• Sources. DataForSEO's AI Keyword Data search volume and Google Ads search volume through DataForSEO, both with twelve months of history, for the United States in English and — ladders and brands only — the United Kingdom. Collected on 14 October 2026 for $0.25.
• Limits. One vendor's AI figure, tested against Google rather than against real AI conversations, which nobody publishes. Google Ads returns no figure for some terms, "vpn" among them, so those phrases drop out of the comparisons. Google's latest month is one behind the AI figure's. The ambiguity cut-off is ours.

Every phrase with its four figures is free to download as CSV. You are welcome to use it with a link back to this page.