Free Query Fan-Out Tool

See how AI search engines decompose a single query into a research strategy. Type any question and watch the fan-out live.

Decomposing your query…
Generating hypothetical answer
1 Pre-Retrieval Prior Before any web search
What the model thinks it knows
This is the model’s hypothetical answer — generated from training data alone, before searching the web. The fan-out queries below are derived from the claims in this answer.
2 Query Decomposition
Broad discovery Specific verification Nice-to-have
3 Retrieval Targets What a perfect result looks like
HyDE Targets

For each sub-query, the model generates a short “ideal answer” — the kind of content a perfect result would contain. Retrieved web pages are then scored by how closely they match these targets.

Expand any sub-query above to see its retrieval target.

How query fan-out works

AI search engines don’t just send your query to a search engine. They run a multi-stage research pipeline before writing a single word.

1

Hypothetical answer

The model drafts an answer from training data alone — no web search yet. This “pre-retrieval prior” anchors everything that follows. It decides what the model already thinks it knows and what it needs to verify.

2

Query decomposition

Claims in the hypothetical answer are analysed and broken into targeted sub-queries — a mix of broad discovery searches (to find things the model might have missed) and specific verification queries (to check the claims it made).

3

Retrieval targets

For each sub-query, the model generates a short “ideal answer” describing what a perfect result would contain. Retrieved web pages are then scored by how closely they match these targets using semantic similarity.

This is what the tool above shows you. But the pipeline doesn’t stop here.

What happens next

The fan-out queries are just the research plan. Here’s what a full AI search pipeline does with them.

4

Web retrieval

Every sub-query is sent to search APIs. Dozens of pages are retrieved, each carrying snippets the model might cite in its final answer.

5

Reranking & selection

Retrieved snippets are embedded and scored against the HyDE targets. Only the most relevant evidence survives — the rest is discarded. This is where your content either makes it into the answer or doesn’t.

6

Grounded response

The model writes its final answer using only the selected evidence. Every claim is grounded in a retrievable source with inline citations.

Answer Engine lets you observe and test this entire pipeline — with persistent persona context, branching conversations, and full source traceability at every stage.

See the full picture

Answer Engine models the person, the conversation, and the retrieval together. That’s why context-free prompt tracking gives you the wrong picture.

Learn more View plans