The knowledge governance problem is what everyone calls GEO

For about eighteen months now the industry has been busy inventing new acronyms for what is essentially the same problem. GEO. AEO. AIO. Brand Sovereignty is the latest entry, and to Bill Hunt's credit it's the sharpest framing I've seen — but even that phrase understates what's actually going on.

The problem is not that your schema is thin. The problem is not that you haven't published enough content. The problem is that most businesses have never treated their own product knowledge as a governed asset, and AI systems are the first distribution surface in history that punishes them for it.

I want to make that case properly, because it changes what the work actually is. If I'm right, then a lot of what's being sold as GEO consulting is treating a symptom two layers up from the disease.

What AI systems are actually doing when they answer

When someone asks an AI system "which mattress is best for a side sleeper who sleeps hot," the model is not ranking pages. It is assembling an answer from whatever structured, semi-structured, and unstructured evidence it can pull confidence from. Product attributes. Reviews. Comparison content. Support documentation. Retailer descriptions. Reddit threads. Its own training memory of the brand.

Most brands are already less authoritative about their own products than the retailers selling them.

The output is a confidence decision, not a relevance one. And the AI will happily reach past the manufacturer's own site to a retailer, a review aggregator, or a forum thread if those sources give it higher-confidence evidence about the specific question being asked.

This is the part nobody wants to sit with. Most brands are already less authoritative about their own products than the retailers selling them. Not because the retailers are better at SEO. Because the retailers have organised their product knowledge around how customers actually decide, and the manufacturer hasn't.

The AI isn't picking sources. It's picking whoever has done the work of organising decision-grade knowledge.

Product data is not decision data

Here is the distinction that matters, and I think it's the one that unlocks the rest of the argument.

Almost every mid-sized business I've worked with in the last decade has product data. Specs, prices, dimensions, materials, availability, warranties, delivery windows. It sits in a PIM, or a spreadsheet, or scattered across a CMS, or (most commonly) in three of those places with slightly different values in each. This data is what feeds product pages, feeds, and schema markup. From a traditional SEO standpoint, it's fine.

What almost none of them have is decision data. The information a customer actually uses to buy. Does it sleep cool. Is it good for side sleepers. Does it handle a partner who moves. Will it work on a slatted base. How does it compare to the direct competitor the customer is already looking at.

Some of that knowledge exists inside the business — in the heads of sales staff, in support tickets, in training materials, in the buying guides someone wrote four years ago and nobody updated. But it isn't organised. It isn't governed. It isn't exposed in any form an AI system can consume with confidence.

So the AI does what it always does. It goes and finds the source that has done the organising. Which is usually a retailer, a comparison site, or Reddit.

Why this looks like an SEO problem but isn't

The reason this keeps getting framed as a GEO tactics problem is that the symptoms show up in AI citations and AI Overview inclusions. So the natural response — and the response most agencies are selling — is to bolt on more markup, add FAQ blocks, restructure H2s, layer in schema types, submit to LLM ingestion endpoints.

None of that is wrong. Some of it helps. But it's downstream of the actual issue.

If your underlying knowledge is incomplete — if you literally do not have a governed, authoritative answer to "which of your mattresses is best for a side sleeper" — no amount of schema will conjure one. You can mark up what you have, and what you have will get cited less than the retailer who has answered the question properly.

This is why I've been sceptical of the whole GEO-as-a-discipline framing since the start, and why most "GEO tools" you've been pitched are effectively rank trackers with different labels. They're measuring symptoms of a knowledge problem the tools cannot solve.

The measurement layer makes this worse

There's a second-order issue that compounds all of this, which is that even if a business does invest in fixing its knowledge base, the measurement instruments to prove the investment worked are broken.

You cannot easily see which AI systems are citing you and which aren't. You cannot see what evidence they're pulling from. You cannot see when a retailer's page displaces you as the source of truth for your own product. The feedback loop that would normally push a business to fix its data is muted, because the signal is buried in AI outputs that don't leave clean tracks in analytics.

So the business optimises what it can see. Traffic. Rankings. Conversion. Meanwhile the knowledge foundation quietly rots — or, more accurately, stays exactly as underdeveloped as it always was, while the world around it moves on.

Data governance is invisible until the day it becomes the only thing that matters.

What the work actually looks like

If you accept the argument, the shape of the work changes materially. It stops being an SEO project with a GEO layer on top and becomes something closer to a knowledge management project with a distribution layer on top.

The questions you'd ask are different. Not "how do we rank for this query" but "do we, as a business, have a defensible, structured answer to this question, and where does it live." Not "how do we add schema to this page" but "what is the canonical internal record of this product's suitability for these use cases, and who owns keeping it current."

That work looks tedious compared to shipping a new content batch. It is tedious. It involves interviewing sales staff, mining support tickets, structuring FAQ content around decision criteria rather than keyword clusters, reconciling contradictory product copy across channels, and building a governance model that keeps the knowledge current as products and positioning change.

It also involves accepting that some of what you thought was product knowledge was actually just product marketing — polished, defensive, and useless to a customer trying to decide between three options.

Where reasonable people will push back

I want to be honest about the limits of this argument, because I've been on the receiving end of the counterpoints and some of them land.

The first pushback is that most small and mid-sized businesses will never do this work. It's expensive, it's slow, and the payoff is diffuse. Fair. I think that's true. The businesses that do it will pull further ahead of the ones that don't, but that's been the pattern for every discipline shift I've watched in eighteen years of this work. The ones who invest in the harder, less legible foundations tend to win. The ones who chase tactics tend to churn agencies and wonder why nothing sticks.

The second pushback is that AI systems will get better at inferring decision-grade answers from thin product data, and the manufacturers who did the governance work will have wasted effort. I don't buy it. AI systems getting better at inference just means they'll get better at inferring from the best available source — which will still be whoever has organised the knowledge, not whoever has the most polished product copy. If anything, better inference makes the manufacturer's disadvantage worse, because the retailer's decision-organised data becomes more usable.

The third pushback, and the most serious one, is that this argument sounds a lot like every "content quality" argument that's been made for a decade, and content quality has been an inconsistent predictor of anything. That's true, but I'd argue what's changed is that the surface being served is now a synthesis rather than a link. When Google returned ten blue links, thin content could hide inside a strong domain. When ChatGPT returns one answer synthesised from three sources, thin content gets displaced by whoever answered the question better. The stakes on knowledge quality just went up.

What to do with this if you're the one making the call

If you run marketing at a business that sells products or services with any real decision complexity, the honest audit question is not "are we doing GEO." It's this: could a smart new hire, given access to your internal systems, write a confident, accurate, decision-grade answer to the top twenty questions your customers actually ask before buying — without having to phone three departments and guess at two of them?

If the answer is no, that's the work. Everything else — schema, markup, content plans, AI visibility tools — is downstream of that.

The businesses that will do well in AI search over the next three years are not the ones with the cleverest GEO tactics. They're the ones who treated their own product knowledge as a governed asset before the rest of the industry realised that's what the game was. The game hasn't actually changed all that much. The consequences of neglecting the foundation have just become visible in a way they weren't before.

That's the loop. And most of the industry is still selling tactics to businesses whose real problem is that nobody inside the company can answer their own customers' questions.