Key Takeaways
- Grounding queries are the searches an AI engine runs for itself before answering. They decide which pages enter the candidate pool — so if you never rank for them, you can never be cited.
- Across 1.19M ChatGPT grounding queries and 612K Google AI Overviews grounding queries observed by GEOly, the two engines behave completely differently: ChatGPT issues ~1 search per answer, Google AI Overviews issues ~8.
- 58.4% of ChatGPT's grounding queries contain the word "best" and 26.9% contain "review" — the engine reaches for listicles and reviews, not your homepage.
- Only 0.1% of ChatGPT's grounding queries are phrased as questions. Optimizing for question-shaped keywords is optimizing for the wrong layer.
- A practical workflow: capture real grounding queries from observation data, cluster them by shape, then build the asset type each shape actually retrieves.
Why prompt research is not keyword research
Most GEO advice tells you to research prompts — the things humans type into ChatGPT. That is the right place to start, but it is not where citation is decided.
Between the human prompt and the final answer sits a retrieval step. The model rewrites the prompt into one or more machine-generated searches, runs them against a search index, and reads whatever comes back. Those machine-generated searches are grounding queries (also called query fan-out). They are the actual demand signal your content has to match.
This creates a gap that trips up most teams. You can rank beautifully for the human prompt and still be invisible, because the engine never searched for the human prompt — it searched for something it wrote itself.
What the data actually shows
GEOly observes grounding queries directly: every tracked answer is captured along with the searches the engine ran to produce it. Here is what 1.8M ChatGPT answers and 77,639 Google AI Overviews observations look like in aggregate.
Metric | ChatGPT | Google AI Overviews |
|---|---|---|