Query rewriting

The Citead team, updated

In short. Query rewriting reformulates a request before retrieval, while reranking reorders the documents retrieved.

Definition: Query rewriting reformulates a request into search queries before retrieval, the step that finds relevant documents. It can correct spelling, add related terms or incorporate conversational context so a search system can match the request to available content.

Key facts

As of October 2026, the documentation describes these implementations:

How does rewriting differ from reranking?

Reranking scores and reorders documents already retrieved. Microsoft's semantic ranking overview describes rewritten queries running on the search engine before semantic ranker rescores the results. Rewriting therefore changes the retrieval input; reranking evaluates the resulting candidates.

The same documentation gives the word "capital" as an example of why meaning matters: its interpretation depends on contexts such as finance or geography. Semantic ranker uses context to assess relevance. Its reranking stage processes only the top 50 results from the initial ranking. This distinction matters when examining why a document enters a result set and how it is subsequently ordered.

How does this apply to web retrieval?

Microsoft's Copilot Studio retrieval guide describes a query optimization step that clarifies meaning, incorporates context from the last 10 conversation turns and produces search-friendly queries. Retrieval then runs against the configured knowledge sources. Public website content comes through Bing's index.

For a website owner, the practical implication is that the question typed in a conversation and the query used to search for content can differ. Rewriting sits within a larger process: Copilot Studio retrieves content, then summarizes it and generates a response. The guide identifies websites being indexed by Bing as a requirement for its public website source.

How can an Azure implementation be checked?

Microsoft's query rewriting guide supplies this request body for its hotels-sample index:

{
  "search": "newer hotel near the water with a great restaurant",
  "semanticConfiguration": "en-semantic-config",
  "queryType": "semantic",
  "queryRewrites": "generative|count-5",
  "queryLanguage": "en-US",
  "debug": "queryRewrites",
  "top": 1
}

It requests up to five rewrites. The debug response exposes them under @search.debug. Microsoft recommends debug for testing and omitting it in production for performance. It also warns that rewrites can omit exact terms, affecting searches that require unique identifiers or product codes.

Can rewriting use configured rules?

Google's Rewriting Queries documentation describes synonyms set by the operator of a Programmable Search Engine. Its example maps the financial search term "interest" to variants including "yield" and "dividend". These variants expand searches automatically.

The same documentation recommends domain-specific acronyms and abbreviations as candidates for synonym rules. This gives a site search operator a concrete way to connect a visitor's wording with specialist terminology. The configuration belongs to that search engine; the documented controls describe its query processing.

FAQ

No. Microsoft's Azure guide describes rewriting as optional. When it is disabled, retrieval uses the original query.

Can a follow-up question use earlier conversation context?

Yes. Microsoft's Copilot Studio guide includes context from the last 10 conversation turns in its query rewriting step.

Sources

  1. Microsoft, "Rewrite queries with semantic ranker in Azure AI Search (preview)", 2 July 2026
  2. Microsoft, "Semantic ranking in Azure AI Search", 5 August 2026
  3. Microsoft, "Enhance AI responses by using Retrieval Augmented Generation", 27 January 2026
  4. Google, "Rewriting Queries", 21 August 2024

Check how a page reads to answer engines. Open the GEO score analyzer

Test a page and see what to fix first

The analysis usually takes less than 10 seconds and needs no account.

400,000+ pages scored in the public benchmark