Prompt volume

Prompt volume is the estimated number of times people ask AI engines a given prompt. It is the GEO-era analog of search volume. But no AI platform counts it. There is no Keyword Planner for ChatGPT. So tools like Profound, Peec, and Otterly guess it from panels and models. Treat every figure as sampled, not counted.

Prompt volume vs search volume

Search volume is counted. The estimate is not. Google's Keyword Planner reads Google's own query logs. It hands you a real demand range. No such log is public for AI engines. OpenAI, Google, and Perplexity do not show how often a prompt is typed. So there is no first-party number to read.

Otterly says it plainly. "There's no Google Search Console-like tool for ChatGPT. Nobody has real prompts at this point." That one fact splits the two metrics. Search volume rests on a census. The guess rests on a sample.

The unit differs too. Otterly's 2025 study found real prompts run 15.1 words on average. Estimated prompts run 8.8. Real prompts carry a personal pronoun 52% of the time. A prompt is a spoken-style sentence. A keyword is a fragment. So even the thing you size is longer and looser than the keyword it replaces.

Search volume is counted from query logs; prompt volume is estimated from panels and models. Every figure is sampled, not counted

How it is estimated, and why estimates differ

Every vendor uses a different source. That is why two tools hand you different numbers for the same prompt. There is no shared ground truth. And no vendor shows its full method.

Profound calls itself "the world's first tool to estimate AI search volume." It draws on double opt-in consumer panels. It tracks observed chats, not API outputs or synthetic guesses. It refreshes weekly across ten countries. Peec surfaces AI-suggested prompts and search volumes next to its tracker. Otterly runs the other way. Its tool expands SEO keywords into prompt ideas. Its own docs call these "estimated prompts, not real prompts" that "lack frequency or volume data."

So you get two methods and two ground truths. Panel data watches a sample of real chats. Keyword data infers demand from Google numbers a model rewrites into prompts. Neither can be checked against a platform's real query stream, because that stream is not public. Search Engine Land's read is the honest one. Prompt-level data is "directional, not definitive." The field is not as mature as keyword tracking became over two decades.

Here is how the three tools this project crawled treat the metric.

Why it matters for GEO

The metric is the closest thing GEO has to a priority lever. You cannot watch every prompt. Across 10,000 prompts, AI share of voice is impossible to track by hand. So you pick the prompts worth tracking. The fair tiebreaker is how much demand sits behind each one. That is what Profound pitches. Rank which prompts to track by the volume behind each topic. Used that way, a guess earns its keep. Read as an exact count, it will mislead you. Size prompts to build the list. Then measure the real outcome, inclusion and citation share, on the prompts you keep.

Related terms

This is one entry in the GEO / AEO glossary of emerging terminology. Nearby terms link straight to it. Query fan-out splits one prompt into many sub-queries. The answer capsule is the block you write to win them. GSVO and AI share of voice track how often an engine names you. Agentic commerce covers AI shopping agents that act on those answers. Grounding and RAG explain how engines pull and cite sources. Zero-click search, semantic triples, and entity SEO round out the glossary definitions. For the full method, read the AI visibility pillar.