Want to Understand How to Optimize for AEO/GEO? Understand How AI Search Works First

GEO or AEO is essentially a marketing challenge that started with, well, marketing. This technology has been presented to us as “answer engines,” and it’s pretty easy to see why. Just read the first piece in this series [USE TEXT TO LINK], and you’ll see pretty quickly that an explanation of the probability-based backbone of this technology is less-than-easy to explain quickly and clearly.

In any case, we now (for better or worse) expect AI to be what we’ve been told it is. And if it’s an answer engine, I should be able to ask it, “who’s leading the AFC East right now?” and get correct information. But therein lies the catch: where timeliness is involved (this week’s standings versus last week’s), I don’t just need information, I need up-to-date information. 

This is why most AI tools rely on a combination of both Model-Native Synthesis and Retrieval-Augmented Generation. Taking that into account, if you’re concerned with visibility for your company, the next thing to understand is the order in which each system leans on one or the other, and how.

How AI Search Works: Mechanics By Platform

I recently used ChatGPT to design a custom winter training program for indoor cycling. When it came time to export the result into individual files that I could add to my Wahoo smart trainer/Zwift set up, I ended up with 19 weeks of workouts ordered alphabetically. That’s not great when a program is supposed to get progressively harder in order to ramp your fitness.

So I asked ChatGPT if a different naming or tagging convention would allow me to order the workouts by week and day. “Of course, let me export the files with a new naming convention, “Week + Workout Number + Workout Type” it replied.

I reuploaded the files, and… no dice. Nothing changed. It turned out ChatGPT had only used Model-Native Synthesis, and the old training data indicated that the naming convention would work. Only after I pushed did it query the web and find Wahoo/Zwift users who recently reported that their old means or ordering workouts no longer worked.

This dynamic illustrates not only that the tool I was using searches the web, but also how important the web search is to synthesizing the right answer.

Account for Efficiency First 

It also introduces the first law of the universe that even AI is beholden to: efficiency. Which is to say that if a tool is Model-Native Synthesis dependent (it uses it first, or exclusively), it’s not going to deploy RAG unless it has to. 

That’s not because it’s dumb or lazy, it’s because all that computation power costs money. Afterall, Sam Altman once famously decried the cost dynamics of users including the words “please” and “thank you” in their prompts. Now extrapolate the expense of analyzing three words versus crawling the internet in real time.

If you’re a legacy brand with a central, core product offering, hanging your hat on Model-Native Synthesis could be enough. Another way of saying that is, if the consensus, across public, online information repositories is that you’re great and your products are great, you could be generally ok. 

But if you’re in an actively competitive space, or you or your products are new or under-exposed, then RAG becomes incredibly important. Otherwise, you’re leaning heavily on model updates to bring you into the mix.

If that feels dicey (and it should) and you’re wondering how to optimize your site for AEO/GEO, make sure to read upcoming dispatches, where we’ll dive headfirst into how these incredible tools crawl the web.

AI search is changing how brands get discovered. Want to know where you stand? Let’s talk!