Microsoft launches Web IQ: Bing-powered grounding API for AI agents
Microsoft has announced Web IQ, a suite of AI-native grounding APIs built on Bing’s search index and designed specifically for how AI agents retrieve information. The announcement, made at Microsoft Build 2026, confirms that Web IQ already powers responses in Copilot and ChatGPT’s web search mode.
Where Bing was built for human searchers, Web IQ is built for AI inference: retrieving the right passages, packaging them as structured evidence, and returning them within the latency budgets that multi-step agentic reasoning requires.
What Web IQ does differently
Traditional search returns ranked documents. Web IQ returns passage-level evidence objects: discrete extracts from relevant pages, selected for their information density and assembled into structured context for the requesting model. This reduces the number of tokens a model needs to process and eliminates the step of parsing full-page HTML.
Microsoft measures retrieval quality using GDSAT (grounding satisfaction), a metric covering completeness, freshness, and authority. GDSAT determines which passages surface in a given response. It is a different evaluation layer from traditional search ranking: a page that ranks well for a query may still have individual passages excluded if they score poorly on freshness or fail to add incremental information to what other sources already provide.
Performance claims: sub-165ms p95 latency across five data centres, and roughly 2.5× faster than the next alternative in comparable configurations.
How Web IQ works
Microsoft rebuilt its retrieval stack from indexing through to orchestration to meet the demands of agentic workloads. Key components:
- DiskANN-based vector search for large-scale approximate nearest-neighbour indexing, extended from Microsoft’s open-sourced embedding research
- Fan-out orchestration layer that interprets an agent’s sub-query, retrieves candidate passages across multiple sources, and ranks and assembles results as structured evidence
- Content understanding models trained specifically for LLM-driven reasoning rather than human relevance signals
- Passage selection that maximises information density while minimising token cost per call
Jordi Ribas, President of Search and AI at Microsoft, described the shift: “Agents want to extract the right information from documents, package it and deliver it quickly.”
Publisher controls
Web IQ inherits Bing’s existing robots.txt compliance and publisher preferences. Microsoft has stated it is engaging with standards bodies, including the IETF, on interoperable frameworks for publisher rights in the AI era.
No new crawler user-agent is introduced by Web IQ. It draws from Bing’s index, so existing BingBot configurations govern what Web IQ can access.
What this means
ChatGPT’s web retrieval is now confirmed as passage-level. Microsoft’s announcement explicitly names ChatGPT as a current Web IQ user. Content that earns inclusion in ChatGPT Search responses is being selected at passage granularity, not page level. Each section needs to be extractable and informative on its own.
GDSAT introduces a named signal framework. Completeness, freshness, and authority are the three dimensions Web IQ evaluates. These align closely with GEO guidance on direct answers, recency, and credible sourcing, but they apply at a passage level within a page rather than at the domain or document level.
No new optimisation actions are required. Web IQ runs on BingBot’s existing crawl. Publisher controls work through robots.txt as usual. The practical implication is a confirmation that passage-quality matters in a way that a page-level view of search ranking does not capture.
Access to Web IQ for third-party developers remains limited: Microsoft is accepting expressions of interest at webiq.microsoft.ai but has not announced general availability or pricing.
Sources
More news
-
Google Analytics Adds AI Assistant as Default Channel Group for Traffic Attribution
Google Analytics 4 now automatically classifies traffic from AI tools like ChatGPT, Claude, and Gemini as its own channel, replacing manual tracking workarounds.
-
Google Search Console's 50-Week Impressions Logging Error Is Fixed
A GSC logging error inflated impression counts for 50 weeks from May 2025. Google fixed it going forward in April 2026, but historical data will not be corrected.
-
GSC Branded Queries Filter Reaches Full Rollout on 11 March 2026
Google Search Console's branded queries filter became available to all eligible sites on 11 March 2026, enabling native separation of branded and non-branded traffic.