OpenAI web search on Bedrock gains domain filtering and user-location controls

AWS announced two additions to the web_search tool available in OpenAI's Responses API on Amazon Bedrock. Developers can now restrict results to an allowlist of domains via the filters parameter, and they can pass a user location to localize results. No new endpoint is required; both are options on existing requests.

Domain filtering is the bigger deal for enterprise builders. Grounded answers are only as trustworthy as their sources. Constraining search to, say, a company's docs, regulators' sites or vetted news outlets cuts hallucinated citations and reduces exposure to SEO spam and poisoned pages, an increasingly relevant prompt-injection vector. Location controls matter for queries such as local regulations, store availability or weather, where generic results mislead.

The announcement also shows how far the OpenAI-on-Bedrock relationship has gone. OpenAI models and tools now appear as first-class options inside AWS's catalog. Web grounding is becoming a competitive line item: DeepSeek just priced its Search and Fetch APIs at $0.75 per 1,000 calls, well below the roughly $4-5 typical elsewhere. Bedrock customers will compare total grounding cost, not just token price.

A practical note: filter parameters must be nested correctly in the request. Misplaced date or domain filters in similar search APIs are often silently ignored, so teams should verify that results actually respect the allowlist. AWS has not detailed pricing changes alongside the feature.