Microsoft continues to expand its IQ suite, a family of services designed to ground AI applications in reliable data. Following the release of Work IQ, Fabric IQ, and Foundry IQ, the company has now introduced Web IQ, a service that brings the power of the Bing search index to AI agents. This new offering is a significant modernization of the retired Bing Search APIs, repurposed for the unique demands of large language models and autonomous agents.
The core idea behind the IQ suite is to transform the data businesses already use into a form that AI systems can leverage effectively. Previous IQ products focused on organizational data, using sparse vector stores to provide grounding information. Web IQ extends this concept to the public web, giving agents access to fresh, general-purpose information beyond the limits of their training data. This is especially important because LLMs are static after training, and their knowledge becomes stale over time. By integrating web search, agents can access current information, compare products, check weather conditions, or track breaking news.
Why agents need web search
Traditional web search is designed for human users. Humans type short queries, scan a list of results, and click through to pages. Agents operate differently. They often need to make multiple searches, refine their queries based on partial results, and synthesize information from many sources. They also require structured, machine-readable output that can be injected directly into a reasoning pipeline. Web IQ is built with these requirements in mind.
Bing's underlying search index is massive, but Microsoft has tuned it to work with the vector-based algorithms used in AI. The service relies on DiskANN, a fast approximate nearest neighbor search algorithm developed by Microsoft Research. DiskANN allows high-speed semantic search without requiring the entire index to be loaded into memory. This design enables Web IQ to scale to billions of documents while maintaining low latency, which is critical for interactive agent workflows.
Agents searching the web often send persistent queries that go far beyond simple keywords. They may need to track changes over time, compare details across multiple sources, or follow a chain of reasoning. Web IQ supports these patterns by returning results that are ranked by semantic similarity to the query, not just by keyword matching. This makes it easier for an agent to find the most relevant information even when the query is ambiguous or paraphrased.
Structured evidence objects for token efficiency
One of the most interesting aspects of Web IQ is its approach to returning results. Instead of delivering whole documents, which would be expensive for LLMs to process, the service extracts relevant passages and packages them into what Microsoft calls structured evidence objects. These objects include the key text, along with citation details such as the page title, URL, and crawl date. This design significantly reduces the number of tokens required for grounding, lowering inference costs and improving response speed.
The focus on token efficiency is a direct response to the changing economics of AI. As LLMs become more powerful, the cost of processing long documents can add up quickly. By sending only the most relevant chunks of text, Web IQ ensures that agents spend their compute budget on reasoning rather than on reading lengthy pages. The service also supports markdown output, which is useful for creating semantic memories that agents can store and reuse later.
Web IQ includes several options for controlling the response format. It can return text, HTML, or markdown, as well as extracted passages specifically selected for context. For video and image searches, the service can generate text descriptions using an LLM if the original content does not include them. This makes it easier for agents to work with multimodal data, even when the source is not fully textual.
Integration with MCP and modern agent frameworks
Web IQ has been designed from the ground up for integration with modern AI tools. It provides an llms.txt file and an OpenAPI description, allowing AI systems to discover its capabilities automatically. More importantly, the service is available through a Model Context Protocol (MCP) server, which is quickly becoming the standard way to connect agents to external tools. The MCP server exposes tools for web search, news, images, and videos, plus a browse option that can fetch content from a specific URL.
This integration means that developers can add Web IQ to existing agents with minimal code. Many popular coding agents, such as those built into GitHub Copilot, already support MCP servers. By configuring Web IQ with a simple JSON file and an API key, developers can start experimenting immediately. The MCP server handles the underlying REST calls, so developers can focus on building the agent logic rather than constructing HTTP requests.
Microsoft has also prepared documentation that is friendly to LLMs. This allows autonomous agents to understand how to use Web IQ without human intervention. The API descriptions include detailed error codes and response structures, helping agents recover from failures and handle edge cases gracefully. This level of attention to machine-readability is still rare in the API world, but it is essential for true autonomous operation.
Grounding beyond your own data
While the IQ suite originally focused on organizational data, Web IQ recognizes that many agents need context from the wider world. An ecommerce agent might use web search to pull competitor prices or product reviews. A supply chain agent might check weather forecasts or shipping delays. A financial agent might need to monitor market news or regulatory announcements. In all these cases, Web IQ can provide the fresh, external grounding data that makes the agent more useful and reliable.
Microsoft has been careful to ensure that Web IQ is used appropriately. The Bing crawler respects standard robots meta tags and focuses on quality over quantity. This means that the data returned by Web IQ is both fresh and sourced from reputable pages. The service also includes information about how recently a page was crawled and how stale the underlying data is, allowing agents to judge the reliability of the information they receive.
It is important to note that Web IQ is not intended for human-facing chat interfaces. Microsoft offers a separate service called Grounding with Bing for chatbots. Web IQ is specifically designed for agent-to-machine interactions, where structured data and fast, cheap responses are more important than visual presentation. This distinction allows Microsoft to tailor the technology to the unique needs of autonomous systems.
The launch of Web IQ signals a broader trend in the AI industry: the move from static training data to dynamic, grounded reasoning. As agents take on more complex tasks, they will rely increasingly on external data sources to inform their decisions. By building a service that combines the scale of a major search index with the efficiency requirements of modern LLMs, Microsoft is positioning itself as a key provider of infrastructure for the next generation of AI applications.
Source:InfoWorld News