Not sure where to start? Take a guided tour.
AI Site Grader
AI Query Scanner
llms.txt GeneratorHow aggressive are AI query rewrites?
Mostly, they are aggressive. Average Jaccard similarity between prompt and search query is only 25.24%, and 31.85% of the 11,521 captured pairs count as complete rewrites.
The key numbers.
- Average similarity
- Average Jaccard similarity between prompt and search query.
- 25.24%
- Complete rewrites
- 3,670 pairs fell into the complete rewrite bucket.
- 31.85%
- Prompt-query pairs
- Captured prompt-to-query transformations.
- 11,521
Models do not just tidy phrasing. They often replace most of the user wording with a more search-shaped query, changing the actual terms used for retrieval.
The evidence.
| Measure | Result |
|---|---|
| Average similarity Average Jaccard similarity between prompt and search query. | 25.24% |
| Complete rewrites 3,670 pairs fell into the complete rewrite bucket. | 31.85% |
| Prompt-query pairs Captured prompt-to-query transformations. | 11,521 |
Population and method
Built from 11,521 captured prompt-to-query pairs observed in OpenAI web search calls, with 100% search-query coverage in the sampled dataset.
Where the evidence stops
This study reflects captured search-query transformations rather than every hidden retrieval step a model may perform.
Some inserted entities reflect category heuristics or market leaders rather than final recommendation outcomes.
The underlying dataset is concentrated in commercial and research prompts rather than all possible conversational traffic.
What to do with this
- Write for the query the model is likely to search, not just the wording a human might type first.
- Add current-year, format, and category framing where it genuinely helps the page answer the query.
- Make sure competitor and category language is present when the model is likely to inject it during retrieval.
Relying on raw user prompts for keyword research is insufficient. Teams must measure and optimize for the post-rewrite queries that models actually use to fetch context.
Questions about this answer.
Cite this answer.
Source study: Query Translation“Mostly, they are aggressive. Average Jaccard similarity between prompt and search query is only 25.24%, and 31.85% of the 11,521 captured pairs count as complete rewrites.”
Mack Grenfell · How aggressive are AI query rewrites?
How AI Translates Your Questions · Trakkr Research · Updated 22 January 2026
Apply the finding
Heavy rewrites make single-query checks fragile. Measure share of voice across a stable prompt library to preserve a useful denominator.
On Google Search, track the AI Overview queries that matter repeatedly rather than assuming one phrasing represents the whole topic.