Synthetic Queries

What are Synthetic Queries?

Synthetic queries are search queries created by a system rather than typed by a person. Typically a language model generates them, either by reformulating and expanding a user’s question into related searches, or by writing questions a document could answer. They are used to broaden retrieval in AI search, to train and evaluate search models, and to simulate buyer prompts.

The term appears in several contexts. In information retrieval research, papers such as Promptagator (Google Research, 2022) used a language model to generate synthetic queries from a few examples to train retrieval models. In search, a Google patent application titled “Search with stateful chat” (US20240289407A1, published August 2024) describes “generating one or more synthetic queries using the LLM output” and selecting documents responsive to both the original and synthetic queries.

How it works

flowchart TD
  A[User question] --> B[LLM interprets intent and context]
  B --> C[Synthetic query 1]
  B --> D[Synthetic query 2]
  B --> E[Synthetic query 3]
  C --> F[Retrieve documents]
  D --> F
  E --> F
  F --> G[Select passages responsive to all queries]
  G --> H[Generate answer with citations]

In an AI search flow:

1. Interpret. The system reads the question and any conversation context.
2. Expand. A model writes related queries: sub-questions, comparisons, constraints and likely follow-ups.
3. Retrieve. Each query runs against the index.
4. Select. Documents are chosen for how well they cover the original and synthetic queries.
5. Answer. The model composes a response from the selected passages.

Google’s public description of AI Mode matches this pattern: in May 2025 it said AI Mode uses query fan-out, “breaking down your question into subtopics and issuing a multitude of queries simultaneously.” Google has not said publicly how closely the patent describes its production system.

From SEO to AEO and GEO

In classic SEO, the query was fixed: the user typed it, and you optimized a page to rank for that string and its close variants. Query rewriting and synonym expansion existed, but the unit of competition was the visible query.

With synthetic queries, the engine searches on the user’s behalf for questions the user never typed. A buyer asking “best SIEM for a mid-size healthcare company” may trigger hidden searches about HIPAA logging, pricing models and integration with an EDR tool. Your page can be retrieved for one of those, even if it would never rank for the original question.

What carries over: topical depth, intent coverage and internal linking across a topic cluster. What changed: you are optimizing for the set of questions around a topic, not one keyword.

What to do about it

1. Map follow-ups. For each core buyer question, list the comparisons, constraints and “how does it work with X” questions that naturally follow.
2. Answer each in its own section or page with a clear heading and direct opening.
3. Cover decision criteria explicitly: compliance, integrations, deployment model, team size.
4. Use simulation tools carefully. Model-generated sub-queries help planning but are not the engine’s actual queries.
5. Expand prompt tracking with variations, while anchoring the list in real buyer language.

Common mistakes

• Optimizing only for the head query.
• Treating simulated fan-out output as ground truth.
• Creating thin pages for every variation instead of strong, well-structured ones.

Related terms

• Query Fan-Out
• AI Mode
• Google AI Mode
• Prompt Tracking
• Agentic Search
• Google AI Overviews optimization

Sources

• Search with stateful chat (US20240289407A1), Google Patents
• AI in Search: Going beyond information to intelligence, Google, May 2025
• AI features and your website, Google Search Central
• Promptagator: Few-shot Dense Retrieval From 8 Examples, Dai et al., 2022