When Microsoft retired its legacy Bing Search APIs and rolled out the modular Microsoft IQ platform, AI architects immediately asked how custom agents were supposed to pull live internet facts without burning thousands of tokens scraping raw HTML. Microsoft Web IQ is Microsoft's dedicated answer—an AI-native web search and grounding suite built specifically for LLMs rather than human browser clicks.
Powering flagship assistants like Microsoft Copilot and ChatGPT at global scale, Web IQ delivers 2.5x faster grounding speed through a single REST and Model Context Protocol (MCP) endpoint at https://api.microsoft.ai/v3. Better yet, it draws a strict isolation boundary between public internet facts and your private corporate estate.
https://api.microsoft.ai/v3) and MCP (https://api.microsoft.ai/v3/mcp).contentFormat=passage extraction returns targeted 3,000-character segments instead of noisy full-page HTML.x-apikey at 15 queries per second.
As we outlined in our Microsoft IQ Stack for Beginners guide, Microsoft splits enterprise intelligence across 4 distinct context layers rather than dumping everything into a single giant index. Each pillar owns a strict data boundary:
Because Web IQ sits squarely at the perimeter, it has 0 access to private company data. Yet through MCP federation inside Foundry IQ, an agent can seamlessly blend live competitor pricing from Web IQ with internal strategy decks from Foundry IQ in a single grounded response.
Traditional search APIs force developers into a painful 2-step scrape loop: fetch a list of 10 blue links, download megabytes of HTML, strip navigation menus, and pray your model's context window does not overflow. According to Microsoft Learn's official Web IQ documentation, Web IQ normalizes pages into structured documents during indexing.
When you call Web Search, News, or Browse with contentFormat=passage, Microsoft's servers slice out only the passages relevant to your query—capped at 3,000 characters per result by default (or up to 500,000 characters in full-document text, HTML, or markdown mode if approved). In Microsoft's 3,000-query blind benchmark, this approach delivered a 79.05 grounding satisfaction score while consuming far fewer input tokens.
For multi-step agent loops, Web IQ also includes a Browse API with liveCrawl: "fallback" and renderDynamicPages: true for JavaScript-heavy URLs missing from the index. Meanwhile, Named Search Scopes let billing accounts enforce reusable allow and deny lists supporting up to 10,000 unique domain or URL selectors—making it simple to restrict competitive research agents to trusted industry sources.
Microsoft recommends starting most grounding workloads with maxResults=10 and maxLength=3000 in passage format, splitting multi-question user prompts into separate focused API calls under the 1,000-character query limit. Notice that Web Search intentionally omits a generic snippet field because the returned content field already carries query-targeted evidence.
Cloud architects planning production rollouts need to understand Web IQ's exact deployment topology before writing Terraform scripts. Even though you deploy a Web IQ resource in a paid Azure subscription (Visual Studio, free trial, and student credits are blocked), that Azure resource only handles billing and metering.
Your application never routes through an Azure VNet to hit it. Instead, every request goes over public HTTPS to https://api.microsoft.ai/v3 via Azure Front Door, using either Microsoft Entra ID OAuth 2.0 tokens (Managed Identity or AgentID) or an x-apikey header bound to your Web IQ AppID.
There are 3 crucial enterprise caveats to note here. First, each AppID starts at a shared 15 queries per second (QPS) rate limit (HTTP 429 above that). Second, Web IQ does not support Private Link, VNet integration, IP allowlisting, or Azure RBAC on the API path. Third, Microsoft provides no formal uptime or latency SLA under the Web IQ Terms of Use—though you can still place Azure API Management in front of your outbound calls for telemetry and governance.
Here is a quick architectural cheat sheet covering Web IQ's endpoints, limits, and governance boundaries:
https://api.microsoft.ai/v3 and https://api.microsoft.ai/v3/mcpx-apikey header over TLS 1.2 / 1.3
If your team builds AI tools for banking or healthcare, pay close attention to Web IQ's privacy and licensing rules. By default, Web IQ logs query telemetry, operates outside the Azure Data Protection Addendum (DPA) and EU Data Boundary, and performs no server-side PII masking—so Microsoft explicitly advises scrubbing sensitive user inputs with Azure AI Language PII Detection before sending queries.
Regulated teams under HIPAA, GDPR, GLBA, or APPI can request 0-query logging through their Microsoft account team, which strips query text, result URLs, and IP addresses from operational logs. On top of that, Web IQ's Terms of Use enforce strict 200-character source excerpt limits, require visible clickable citations for every source your model uses, and strictly forbid caching raw search results into a derivative database.
In practice, that means your application can store generated LLM summaries inside per-user chat histories or internal memos, and hold raw content briefly during transient "Atomic Action" agent steps, as long as it deletes raw search payloads immediately once the workflow finishes.
Microsoft Web IQ cleanly fills the public-web gap in the Microsoft IQ ecosystem, pairing Bing's global index with token-efficient passage extraction and plug-and-play MCP support. Just remember to budget for its public HTTPS architecture, 15 QPS starting cap, and strict non-caching terms before going live. Are you pairing Web IQ with Foundry IQ or Work IQ in your enterprise agents? Let me know in the comments below!