How people actually ask AI about products and services
A map of real human prompts, not invented examples and not keyword tools. Every number here comes from 2,909,549 deduplicated prompts people really typed.
What people use AI for
Most prompting is not commercial. Sorting that out first is what stops a share-of-voice number being quoted against the wrong denominator.
| Intent | Prompts | Share | |
|---|---|---|---|
| 1 | General information | 1,826,704 | 42.5% |
| 2 | Other or unclassified | 1,722,733 | 40.0% |
| 3 | Commercial | 660,387 | 15.4% |
| 4 | Coding | 76,832 | 1.8% |
| 5 | Personal | 14,862 | 0.3% |
| 6 | Business shopping | 660 | 0.0% |
Commercial intent is 15.4% of cleared prompts. That is the pool worth optimising for.
The commercial prompt types, ranked
Within commercial prompting, this is what people are actually asking. Read it as the jobs AI is being hired to do when someone is shopping.
| Commercial prompt type | Prompts | Share | |
|---|---|---|---|
| 1 | Which brand should I pick | 200,858 | 27.7% |
| 2 | Tell me about this category | 154,760 | 21.4% |
| 3 | What does it cost | 111,394 | 15.4% |
| 4 | Find me someone local | 73,284 | 10.1% |
| 5 | Compare these two | 62,350 | 8.6% |
| 6 | Is it any good | 49,513 | 6.8% |
| 7 | I am ready to buy | 48,564 | 6.7% |
| 8 | What else is there | 15,273 | 2.1% |
| 9 | Can I trust them | 8,796 | 1.2% |
Where they are in the journey
| Stage | Prompts | Share | |
|---|---|---|---|
| 1 | Awareness and research | 350,279 | 48.9% |
| 2 | Purchase decision | 157,620 | 22.0% |
| 3 | Consideration and comparison | 77,475 | 10.8% |
| 4 | Local action | 73,284 | 10.2% |
| 5 | Validation and trust | 58,185 | 8.1% |
A prompt can carry more than one stage, so these rows sum to more than the commercial total. Stage is assigned per prompt, not derived from the type above.
How the ask is framed
| Shape | Prompts | Share | |
|---|---|---|---|
| 1 | Recommendation | 198,970 | 39.6% |
| 2 | Budget-bounded | 75,802 | 15.1% |
| 3 | Local-service | 73,284 | 14.6% |
| 4 | Comparison | 62,350 | 12.4% |
| 5 | Brand-validation | 47,035 | 9.4% |
| 6 | Feature/integration constrained | 18,835 | 3.8% |
| 7 | Alternatives | 14,909 | 3.0% |
| 8 | Shortlist | 9,474 | 1.9% |
| 9 | Persona/use-case constrained | 1,199 | 0.2% |
What they constrain the answer by
The constraint someone attaches to a prompt is the thing your page has to answer. Feature and price dominate, and location matters far more than most content plans assume.
| Qualifier | Prompts | Share | |
|---|---|---|---|
| 1 | Feature | 304,697 | 49.4% |
| 2 | Price | 96,881 | 15.7% |
| 3 | Budget | 75,802 | 12.3% |
| 4 | Location | 73,284 | 11.9% |
| 5 | Brand | 56,745 | 9.2% |
| 6 | Integration | 8,380 | 1.4% |
| 7 | Persona | 1,199 | 0.2% |
Short questions and long consultations are different behaviours
Averaging these together is how you get a meaningless "average prompt length". They are two distinct modes and they need different content.
| Length band | Prompts | Share | |
|---|---|---|---|
| 1 | 1 to 5 words | 735,912 | 17.1% |
| 2 | 6 to 10 words | 449,041 | 10.4% |
| 3 | 11 to 25 words | 611,405 | 14.2% |
| 4 | 26 to 60 words | 447,998 | 10.4% |
| 5 | Over 60 words | 2,057,822 | 47.8% |
Published reports
Each report is the same analysis narrowed to one industry, with real example prompts and a prioritised build list.
Run this yourself, in Claude
The corpus is hosted privately on Cloudflare R2. The skill pulls it once (about 1.3 GB, roughly two minutes) then runs locally, so every run after the first is fast and costs nothing.
Three steps
- Install the skill folder, and rclone if you do not have it.
- Drop the read-only R2 credentials at
~/.r2-corpus.conf. Ask Joe for them. - Run
python scripts/rocket_geo_intelligence.py run --industry <name>. The corpus downloads itself on first use.
Adding a new industry is one JSON config in configs/. Copy
_industry_template.json, set the terms, run it, publish.
Read the limits before you quote a number
These are prompt-phrasing counts from a fixed corpus. They are not search volume, not market size, not demand forecasts and not AI answer share of voice. Nothing here tells you how often a brand is cited in an AI answer. Percentages are shares of this corpus. Small counts are labelled directional and should be verified before they drive spend.
Corpus lanes are kept separate on purpose. Public LLM-chat data is not confirmed ChatGPT traffic, and traditional-query data is not AI-chat behaviour.