Microsoft Web IQ: Why keywords are the last line of defence in the agentic web

Microsoft recently published a look under the bonnet of Web IQ, their new grounding platform for AI agents. The piece — "Grounding at scale: engineering the retrieval system for the agentic web" — outlines a bold architectural shift. Instead of traditional search pipelines, Web IQ runs on a semantic-first architecture: it indexes content down to the passage level using vector embeddings, and pushes keyword matching to the very end of the pipeline to "refine precision where necessary."

It sounds like a flawless, next-generation evolution of search. But I'm not convinced.

While a semantic-first approach looks great on paper, it ignores a massive, looming reality the whole industry is currently pretending isn't happening: the explosion of AI-generated content. We're flooding the web with billions of synthetic documents, and in that context, demoting keyword matching to a backseat role feels less like evolution and more like wishful thinking.

Here's my opinion on why keywords aren't just a legacy fallback but rather the last line of defence against the sludge.

Vector embeddings struggle in a noisy world

Embeddings are genuinely brilliant at intent, nuance, and abstraction. Search for "troubleshooting a leaky kitchen pipe" and a vector model will happily connect you to a passage about "fixing a dripping sink" without a single word in common. That's the magic.

But vectors are bad at exactness — and that weakness turns into a liability the moment the data pool gets noisy. When a system leans purely on the geometric vibe of a text to judge relevance, it's quietly assuming the underlying corpus is distinct and varied. The modern web is becoming the opposite: a homogenised soup of AI-generated articles, trained on the same data, written in the same corporate syntax, optimising for the same algorithmic visibility.

The threat of "Semantic Flattening"

As billions of synthetic documents inflate the grounding territory, we run into semantic flattening, in other words the practice of grouping different entities under a single umbrella, erasing those important distinctions and nuances in the meaning.

LLMs generate text from statistical probabilities, so AI-generated content tends to sound remarkably alike — same sentence structures, same writing style, same vocabulary. When an embedding model maps those pages into geometric space, thousands of low-quality content-farm articles end up clustering nearby to the authoritative human sources.

That's the blind spot. To a semantic encoder, a carefully written 2024 academic study and a 2026 AI-generated blog post summarising that same study look nearly identical in intent. If pure vector search is doing the heavy lifting of pulling the initial candidate set, the agent is liable to drown in a sea of synthetic echoes long before it ever reaches the source truth.

I've seen a smaller version of this problem these days while building my own tool. Frequency is a terrible proxy for importance — and so, it turns out, is geometric proximity once the space gets crowded.

Keywords aren't retiring — their usage is being re-purposed

This is why I don't see keyword matching disappearing anytime soon. If anything, the AI inflation of the web makes lexical matching more critical than ever.

Microsoft frames Web IQ as shifting keyword matching to the later stages of the pipeline. Let's call that what it actually is: a compliance check.

When the web is polluted with synthetic text, keywords become an unyielding anchor to reality. If a user — or an agent — needs a specific part number (277-11076-ND, a digiKey), a precise legal reference (art. 615-ter c.p. of Italian law, unauthorised access to an IT system), or an exact chemical compound, a vector model will surface hundreds of relevant-sounding articles that hallucinate or generalise the detail. A later-stage keyword pass is the only thing standing between the accuracy of the answer and a confident fabrication. It's a strict gatekeeper: "does this passage actually contain the exact identifier? No? Bin it."

Conclusion: the grounding reality check

Building a semantic-first architecture for AI agents is a serious engineering achievement, and Microsoft's passage-level indexing is almost certainly where things are heading. But a retrieval system has to be designed for the web we actually have, not for a utopian system that will never exist, or in fact that we are distorting with our way of working. Vectors understand the vibe of what you're looking for, but until we solve the synthetic-pollution problem, the old-fashioned keyword matching is the sole thing keeping these agents anchored to reality.