This is a nice, neat summary of the core constraints of current LLM based AI when it comes to SEO/GEO (based on a much longer, more technical piece, if you want the details).
Back when ChatGPT 3.5 came out, I was telling anyone who’d listen that it was going to disrupt search and publishing.
In early 2024, while at PwC, I started pitching new content formats to address this – intended to help capture whatever the GenAI equivalent of search ranking was going to be. “GEO” before this label stuck (I was calling it AIO at the time).
My thinking then was based on what seemed to be a logical, structured approach – similar to the “query fan out” advocates you’ll see in the “GEO” space today. (Basically label the hell out of your content, anticipate and answer the questions your target audience is likely to ask, as that structure should help the AI understand the context more easily, and so encourage it to pull from your page rather than someone else’s. Effectively a slightly deeper version of an old school Q&A or FAQ piece…)
But as I dug deeper it soon became clear that the challenge with LLM-based GenAI (from a model visibility perspective) wasn’t to do with clarifying the intended meaning of the information you want the model to ingest and regurgitate, as I first thought. (“These things can process unstructured data, but they’ll process *structured* data easier – so let’s structure it for them.”)
Instead it’s that these systems – despite being called Large *Language* Models – don’t actually understand language, or context. “Logic” to them is a meaningless concept; not only that, they have no concept of what a concept even is.
—
Tokens aren’t words, and don’t have meaning independently – they only appear to have meaning when combined into words.
Tokens create the illusion of being words (and having meaning) because of the probabilistic nature of these tools, when working with them using language as the system interface. This creates an environment in which they’re working within the rules of language, so can produce output that makes sense – even if they don’t “understand” what they’re saying.
But URLs aren’t language, and don’t have linguistic rules or any consistency from site to site in terms of information architecture. Every site’s URL structure is similar, but different.
And as LLMs don’t really understand structure (except as recognisable, predictable patterns), this makes accurately relating URLs a significant challenge for current LLM-based GenAI tools.
—
This is a structural challenge, baked into the very nature of these models. Despite what many GEO “experts” are now claiming, if your goal is to generate links and traffic from GenAI results, it’s not going to be an easy one to engineer if you’re working from outside that system.
It may be possible to tweak model outputs to improve this and increase URL attribution accuracy, but a) it won’t remove the underlying structural constraints, and b) what would be the incentive for the GenAI companies to do this?
The dust has yet to settle on this one.
