Bing just told the industry what GEO actually is

Microsoft's Bing team published something yesterday that should have been the bigger story of the day. Buried under the Google AI Overviews link updates and the OpenAI ads manager rollout, the Bing engineering team posted a framework laying out how grounding indexing differs from search indexing — and in doing so, quietly described what most of the GEO industry has been failing to articulate for two years.

It's the clearest engineering-side statement we've had on what AI search systems actually need from the web. And if you read it carefully, it confirms a position I've been arguing for months: GEO isn't a separate discipline. It's traditional SEO with the tolerance for error stripped out.

The whole post is worth reading. The thing that matters is what it implies about how publishers should be thinking, and almost nobody covering it has connected those dots.

What Microsoft actually said

The Bing team's argument is that traditional search and grounding share infrastructure but optimise for different questions. Search asks "which page should this user visit?" Grounding asks "what information can the AI safely use to construct a response?"

That distinction sounds academic. It isn't. It cascades through every part of how the index has to behave.

Microsoft lists five places where requirements diverge. Factual fidelity matters more, because there's no human in the loop to catch a misranked result — the AI just speaks. Source attribution becomes a core signal rather than a nice-to-have. Freshness costs more, because a stale fact in a generated answer is misleading in a way a stale ranking isn't. Coverage of high-value facts has to be near-complete, because there's no fallback. And contradictions can't be silently arbitrated, because an AI confidently picking the wrong source is worse than a search results page presenting both.

Then they add two design choices. Abstention — the system declining to answer when evidence is weak. And iterative retrieval — the system asking follow-up questions and refining its search mid-response.

This is the engineering specification GEO should have been written from.

What the GEO industry has been telling people instead

The GEO playbook being sold to UK businesses right now mostly consists of: add more schema, write FAQ blocks, structure content for snippet extraction, optimise for "AI-friendly" formatting. There are tools charging four-figure monthly subscriptions to monitor citation share. There are agencies repackaging traditional content audits as "GEO readiness assessments."

Read against the Bing framework, almost none of that addresses what the systems actually need.

What the systems need, according to the team building one of them, is: factually accurate information, well-attributed, fresh enough to not mislead, comprehensive on the high-stakes facts, and consistent across your own surfaces. That's not a GEO checklist. That's editorial standards. It's the same thing serious publishers have been doing for a hundred years.

Schema helps because it makes attribution machine-readable. Structured content helps because it survives chunking. But these are second-order tactics. The first-order question is whether your content is the kind of thing a grounding system would responsibly cite, and that question is answered by editorial discipline, not markup.

Abstention is the part nobody is talking about

The single most important word in the Bing post is "abstention." It describes a grounding system declining to answer because the evidence is missing, stale, or contradictory.

Think about what that means for publishers and brands. If the system can't find a confident, well-attributed, current answer about your industry from sources it trusts, it doesn't make one up — increasingly, it just doesn't answer. Or it answers with whatever weak source happens to be available, which is often not you.

The mental model most people still have is that AI search ranks content the way Google ranks pages, just in a different format. That's wrong. AI search is constantly making a binary decision about whether to cite anything at all on a given query. If your sector has thin coverage, weak attribution, or contradictory sources, the system might just stay quiet — and the visibility you'd have got from a normal SERP listing simply doesn't exist.

That's a category of loss the SEO industry isn't measuring because it can't measure it. You can't see a query that didn't trigger a citation. You can't track a brand mention that didn't happen because the system abstained.

Iterative retrieval breaks the single-page mindset

The other design difference Microsoft flags is iterative retrieval. The system asks follow-ups, refines, combines evidence across multiple sources before generating a response.

The ranking unit isn't the page anymore. It's the fact.

This matters because it kills the page-as-unit thinking that SEO has been built on for two decades. The ranking unit isn't the page anymore. It's the fact. And the fact has to be consistent across every place on your site it appears, because the system might pull it from any of them — or worse, pull contradictory versions from two of them and then abstain.

Most websites I audit can't pass this test. Pricing on a service page doesn't match the pricing on the FAQ. The "about us" date doesn't match the founder's LinkedIn. The product spec on the category page contradicts the spec on the product page itself. In a traditional ranking system, this is a minor quality signal at worst. In a grounding system, it's a reason to discard you as a source entirely.

After 18 years of this work, the pattern is depressingly consistent. The sites that win in AI search are the sites that already had editorial discipline before AI search existed. The ones losing are the ones that treated their website as a marketing brochure where minor inconsistencies didn't matter.

What this means if you actually run a business

Three implications, which I'd argue matter more than anything in the GEO tool catalogues:

The first is that consistency across your own surfaces is now a ranking factor in a way it wasn't before. Every page that mentions your pricing, your hours, your team, your service area, your specs — they all need to agree. Not approximately. Exactly. This is unglamorous work and most agencies won't propose it because it doesn't sound like SEO.

The second is that your content needs to be confidently citable on the specific facts people ask about your sector. Not comprehensively — confidently. Microsoft's framework specifically calls out coverage of high-value facts as a grounding requirement. If the most-asked questions about your category aren't answered clearly and authoritatively on your site, you're not in the running, regardless of how much schema you've added.

The third is that the freshness bar has moved. Stale content used to be a slow ranking decay. In grounding systems it's a binary — either the fact is current enough to cite or the system finds someone else. The "set and forget" content model is dead for any topic where facts change.

None of this requires a new tool. None of it requires a GEO platform subscription. It requires the kind of editorial care that good publishers have always practised and that most marketing-led websites have always cut corners on.

That's the loop. And we built it.

The honest bit

The Bing post is one company describing one system. Google's grounding architecture for AI Overviews and AI Mode may differ in important ways. ChatGPT's retrieval pipeline definitely does. The five categories Microsoft names are useful as a framework, not gospel.

But the underlying point — that grounding has lower tolerance for error than ranking, and therefore demands more from publishers — is structural, not specific to Bing. Any system that generates direct answers from web content faces the same constraints. Any vendor telling you their GEO methodology is fundamentally different from good SEO is selling you something.

The most useful thing the Bing team did this week wasn't announce a feature. It was tell the industry what's actually being measured. The fact that almost nobody covered it that way says more about the state of GEO commentary than it does about Microsoft.