The bifurcated web won't last

Everyone's optimizing websites for AEO based on what LLMs do now. But a lot of current AEO tactics seem to treat the bifurcation of the internet (humans consuming one version, AI search systems consuming another) as something that will endure.

I'm not so sure. Given how quickly models are advancing, particularly in image, video, and voice, a future seems imminent where AI systems can understand a website without needing its context and content served up primarily as text, pre-rendered HTML, or markdown.

I'd argue that our current reality feels more like an implementation constraint (perhaps due to compute, cost, latency, reliability, etc.) than a fundamental limitation of what models can understand.

If multimodal models can increasingly understand images, video, interfaces, and spatial relationships, why should we assume AI search will remain disproportionately dependent on extracting text from a page?

I'm seeing companies sacrifice design sensibilities to ensure AI search inclusion. It feels a little like the new Wild West of keyword stuffing.

It all makes sense given incentives (my favorite topic from my Econ days): as AI search becomes an increasingly important discovery channel, nobody wants to be invisible in the conversations where their buyers are doing research.

And it's true that today's LLM crawlers like having things explicitly spelled out, if not visibly on the page, then through structured content and schema. That makes current AEO tactics a rational response to the limitations of the technology, even if they often feel kind of icky. (Very scientific, I know.)

Great design, on the other hand, leaves a lot explicitly unsaid. It makes you feel something and allows you to deduce a brand's identity not just from words, but from typography, spacing, imagery, among other more experiential factors.

So I get the tradeoffs companies and marketing teams within them are faced with, even if I don't like the output in many cases.

I just think that ultimately, AI search systems and human web visitors are trying to get to the same things: clarity and comprehension. It's just that the cues humans use to that end are much more nuanced.

My hope is that as the constraints lessen, models get better at interpreting the same visual and contextual signals we do, so that the LLM version and the human version of the web start to converge.

Looking forward to the day we no longer need to plaster FAQs on the homepage, and can instead assume that by building to delight and inform humans, we'll also do a good job of educating our AI sidekicks.