Microsoft Web IQ Claims 164ms Grounding Latency—What Developers Should Test

What Microsoft is claiming​

Web IQ is Microsoft's suite of AI-native web search and grounding APIs for language models, agents and applications. Microsoft announced it on June 2, 2026. The announcement came in a Bing Search Blog post by Knut Risvik, Distinguished Engineer for Search and AI. It describes Web IQ as a suite built for the agentic era, drawing on web pages, news, images and videos.

A newer Microsoft engineering post, "Why AI agents need external grounding," restates the performance case. The headline numbers:

• Latency: P95 latency measured 164 milliseconds across five datacenter regions, against 406 for the closest competitor and 2,090 for the slowest. The "2.5x" figure is simply 406 divided by 164.
• Quality: Web IQ scored 79.05 on "grounding satisfaction" in a blind test of 3,000 queries sampled from production traffic. The next-best service scored 75.70. Microsoft says this is a benchmark score and not an accuracy percentage.
• Token use: Web IQ reportedly reached peak quality at about 40,000 tokens per query. Competing services used 60,000 to 100,000 tokens and scored lower.

Microsoft's FAQ says the comparison used 10 results and 10,000 characters per result, or the equivalent. It also says results depend on the query set, the model and the evaluation method.

Why grounding matters to agents​

The engineering post argues that a model's training data goes stale and is too broad for specialized questions. Examples are product versions, vendor advisories and local regulations. Retrieval at answer time puts current evidence into the model's context and gives users source links to check.

Microsoft is careful to say grounding does not guarantee correctness. It gives people a better basis for checking an answer before they act on it.

An agent also differs from a person using a search engine. A person can skim results, discard weak pages and rephrase the query. An agent may consume the retrieved passages directly and act on them. That raises the stakes for freshness, relevance and attribution.

How to read the numbers​

The comparison is Microsoft's own. The competitors are not named, and the published text doesn't give enough detail to reproduce the test. Another outlet, WindowsForum's earlier coverage, made the same point: outsiders can't fully audit the comparison set, so it is best read as a company claim and not an independent ranking.

P95 retrieval latency is not end-to-end response time. P95 means 95% of calls finished at or under that time. It measures the retrieval call only. Microsoft does note that latency compounds in agent workloads, because a multi-step chain pays the retrieval cost at every hop. But a deployed agent also spends time on model reasoning, extra tool calls and orchestration. A third-party Foundry workshop makes the same distinction. It lists the Web IQ SDK at 164 ms p95 for search only, and a Responses API web search at roughly 2 to 5 seconds because it combines search and model time. Those workshop figures are not Microsoft benchmark data and are only a rough guide.

Grounding satisfaction is not accuracy. A 79.05 does not mean 79% of answers are correct. The 3.35-point gap over the runner-up is also modest next to the latency gap.

Token savings depend on configuration. Microsoft credits passage-level retrieval, meaning it prioritizes relevant passages and cuts unnecessary context. Your savings will depend on your queries and how you set result counts and content formats.

Evaluating a grounding service​

Microsoft itself recommends testing against your own workflow. Its suggested areas, with my practical notes:

1. Freshness. Test queries about recent changes. Separate when a source was published or updated from when the service fetched it. Different workflows tolerate different delays.
2. Coverage. Check that the service finds the sources and specific details you need. Include niche terms, product versions, local information and hard-to-find documentation.
3. Total task cost. Count retrieval calls, model input and output, retries, orchestration and maintenance. Compare services at a similar quality level and at your expected production volume.
4. Citation quality. Check that links and metadata work, that passages keep important qualifications, and that each citation supports the claim attached to it.
5. Governance and operations. Decide who owns retrieval failures, freshness checks and monitoring. Review authentication, data handling, content-use terms and what actions the agent may take.

A sensible test plan is to run the same few hundred real queries through each candidate. Record p50 and p95 latency from your own network location, and compare total tokens consumed at the point where answer quality stops improving. Have someone check citations by hand on a sample.

Availability and product context​

Web IQ is not generally available. It is in limited access by waitlist, and pricing is not announced yet, according to one third-party writeup. Microsoft's FAQ says access goes to select enterprise customers building AI at scale, prioritizing those who work with Microsoft account teams.

Microsoft's FAQ says integration works over REST, MCP via JSON-RPC 2.0, or an SDK. It says the service returns structured JSON with titles, URLs, snippets, timestamps and provenance. It also says Web IQ is distinct from Grounding with Bing, which stays available for existing customers.

The FAQ notes that Web IQ is tuned for LLM consumption and not for human browsing. If you see fewer results than expected, it suggests several checks:

• Inspect the raw response.
• Verify the maxResults setting.
• Check content eligibility and SafeSearch behavior.
• Check any site: or -site: operators in the query.
• Check whether a client-side or downstream step is filtering results.

It also says the index is continuously refreshed and updates propagate across data centers. Some variation between identical requests is therefore expected.

Web IQ sits alongside three other parts of Microsoft's IQ family. Microsoft describes it as a suite that gives LLMs, agents and applications access to current, relevant web information. Microsoft Learn describes the others this way:

• Work IQ: context on people, collaboration and workflows.
• Fabric IQ: the live state of the business, including entities and rules.
• Foundry IQ: curated institutional knowledge such as policies and authoritative documents.

Web IQ supplies the public-web layer in that stack.

Bottom line​

The 164 ms versus 406 ms comparison is a real, specific claim, but it is Microsoft's own measurement of retrieval latency only. It makes Web IQ worth a test if you build agents on Microsoft's stack and are waiting on the access list. It does not show that finished agent responses will be 2.5 times faster, or that answers will be more accurate in your domain. Measure your own end-to-end latency, cost per task and citation quality before you commit.