What is a synthetic query?
A synthetic query is a search query that an AI model writes, rather than one a person types. If someone asks Google's AI Mode "is it worth repairing a 10-year-old dishwasher", the system may also search for things the person never typed, such as typical dishwasher lifespan, common repair costs and energy use of new models. Those generated searches are synthetic queries.
The term is used in two related ways:
- In AI search, synthetic queries are the extra searches an AI system runs during query fan-out to gather material for an answer.
- In search research, synthetic queries are machine-written questions used as training data, so a retrieval system can learn which documents answer which questions without needing millions of real, labeled searches.
Both uses share one idea: a language model guesses the questions a document answers, or the questions a user really needs answered, and those guesses are then used to retrieve documents.
How do synthetic queries work in AI search?
In AI search, synthetic queries work by breaking one user question into many smaller searches, running them at the same time, and combining what comes back. Google described this in March 2025 when it introduced AI Mode: the system issues "multiple related searches concurrently across subtopics and multiple data sources and then brings those results together" [2]. Google's guidance for site owners defines fan-out as "a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results" [6].
The generated queries are synthetic queries. In May 2025, Google said AI Mode issues "a multitude of queries simultaneously on your behalf", and that Deep Search, a heavier research mode, "can issue hundreds of searches" for a single question [3].
The practical effect is that a page can be retrieved for a question it was never written to rank for, because it matched one of the hidden synthetic queries rather than the user's words.
What types of synthetic queries does Google generate?
Google has not published a list of synthetic query types. The most cited breakdown comes from Mike King of iPullRank, who reads Google patent filings and identifies seven kinds [4]:
| Type (per iPullRank's reading of patents) | What it does |
|---|---|
| Related | Searches for adjacent topics |
| Implicit | Searches for needs the user did not state |
| Comparative | Searches comparing products or options |
| Recent | Draws on the user's earlier searches |
| Personalized | Shapes searches to the user's interests |
| Reformulation | Rewords the original question |
| Entity-expanded | Swaps in related entities, such as other brands |
This list is an analyst's interpretation of patents, not Google documentation. Patents describe methods a company has claimed, and a filing does not prove the method runs in a live product.
When were synthetic queries introduced?
Synthetic queries were used in search research years before AI search products: Google researchers showed in 2022 that a language model could generate training queries from just eight examples. Google then brought model-generated queries into consumer search with AI Mode's query fan-out, first described on March 5, 2025 [1] [2].
2022: synthetic queries as training data
In September 2022, Google researchers published Promptagator, which used a large language model "as a few-shot query generator" to create task-specific training data for retrieval systems [1]. With no more than eight examples, the generated queries trained retrievers that outperformed heavily engineered models trained on large labeled datasets [1].
2025: synthetic queries in Google Search
On March 5, 2025, Google introduced AI Mode and named the "query fan-out" technique [2]. On May 20, 2025, Google rolled AI Mode out to everyone in the US and described Deep Search's ability to issue hundreds of searches [3]. Several SEO guides date fan-out to the May 2025 announcement; Google's own description appeared two and a half months earlier.
What is the difference between a synthetic query and query fan-out?
Query fan-out is the process, while synthetic queries are what the process produces. Fan-out describes an AI system splitting one question into many searches. Each of those searches is a synthetic query. A related distinction: a query refinement is a new search the user types themselves after seeing results, while a synthetic query is generated by the system without the user seeing it.
What is known and what is claimed about synthetic queries?
What Google has confirmed is limited: its AI search features issue multiple model-generated searches for one question, and Deep Search can issue hundreds [2] [3] [6]. Everything else is inference.
| Statement | Status |
|---|---|
| AI Mode runs multiple related searches per question | Confirmed by Google [2] [3] |
| Deep Search can issue hundreds of searches | Confirmed by Google [3] |
| Synthetic queries are matched to passages using vector embeddings | Described in patents and analysis [4] [5], not confirmed for live products |
| Specific query types (implicit, comparative, entity-expanded and so on) | Analyst reading of patents [4] |
| How many synthetic queries an average answer uses | No published figure |
| Which synthetic queries were run for a given answer | Not shown to users or site owners |
Third-party tools that claim to show "the fan-out queries" for a keyword generate their own guesses with a language model. They can be a useful brainstorm, but they are not Google's actual queries.
How do you optimize for synthetic queries?
You optimize for synthetic queries by covering the questions that naturally sit around your main topic, so your pages can answer the sub-questions an AI system is likely to generate. For a page about repairing dishwashers, that means also answering lifespan, cost, replacement versus repair and common faults, each clearly, in its own section.
Google's guidance warns against overdoing it: site owners do not need to chase every long-tail variation of a query, because its systems understand synonyms and general meaning [6]. Write sections that each answer one clear question in the first sentence, so any single section can be retrieved and used on its own.