Brand visibility is now a Bayesian problem

AI systems don't retrieve your content. They reason about your category using everything they absorbed during training, and your position in that answer depends almost entirely on what they absorbed before the conversation started.

How AI systems actually form a view of your brand

In Bayesian inference, you start with a prior, a belief formed from everything you've seen before, and update it as new evidence arrives. Large language models work the same way. They have priors baked into their weights from training: associations between companies, categories, claims, and credibility signals. When someone asks a question, the model integrates those priors with whatever is in the current context to construct an answer.

Your brand's position in that answer depends almost entirely on the prior. The prior is built from training data, and training data skews heavily toward content that is public, cited, discussed by third parties, and structurally clear. Not SEO-optimised landing pages. Not gated whitepapers. Not thin category copy written for a crawler.

If the prior is weak, no amount of content production will rescue it. You're fighting the weights.

Why SEO is the wrong analogy

For twenty years, B2B marketing treated search visibility as a content volume and authority game. Publish more, earn more links, climb the rankings. That model worked because the mechanism was indexing: Google read your page and filed it.

Generative Engine Optimisation (GEO) is a different mechanism entirely. The term was coined by Aggarwal et al. in a 2023 paper from Princeton, Georgia Tech, and the Allen Institute for AI. Their framing was about optimising content for AI-generated responses rather than ranked lists. But the deeper implication is that the optimisation target is the model's prior, not the model's retrieval.

That's a harder problem. Updating a search ranking takes weeks. Updating a model's parametric memory takes months of sustained signal: third-party citations, structured content in places models draw from, consistent association between your brand and the specific problems you solve.

Answer Engine Optimisation (AEO) is the shorter-term complement, the tactical layer. FAQs, schema markup, structured comparisons: formats that give models clear, citable evidence when constructing answers. AEO can move faster. But it works best when the prior is already favourable, because a model with a weak prior about your brand won't reach for your structured content when it could reach for a category leader's instead.

The practical implication is that most B2B companies are doing the visible tactical work while neglecting the thing that determines whether any of it compounds.

The prior is a brand problem

The prior is built from everything the domain says about you, not everything you say about yourself. That's the part most B2B marketers haven't fully internalised.

Models don't weight your own claims especially highly. They weight external corroboration: analyst coverage, forum discussions, academic citations, peer reviews, third-party comparisons. The signals that built a strong brand in the pre-AI era, recognition, reputation, consistent association with a problem, are exactly the signals that build a strong prior in an AI system.

Brand and AI visibility aren't parallel problems. They're the same problem, at different timescales.

A useful comparison comes from consumer behaviour research, where Bayesian models of brand evaluation describe how buyers revise brand expectations as they encounter new information. AI systems and human buyers are, structurally, doing the same thing. Both start with a prior, both update on evidence, and both are more likely to reach for brands with strong, consistent priors when the stakes of being wrong are low and the cost of searching for alternatives is high.

The implication for B2B marketers is uncomfortable. Visibility in AI systems isn't a content strategy problem you can solve next quarter. It's a brand investment problem with a long compounding curve.

The companies building that prior now will be harder to displace in two years than any number of well-structured FAQ pages will make you.

---

Related reading

• I am invisible on every AI platform. Here is the data.
• What a 2019 NLP paper tells us about getting recalled by AI