The 3 AI Behaviors That Make AEO Strategy Harder

When it comes to AEO strategy, we don’t yet have prevailing wisdom from which to pull best practices. Those are still being formed. And it makes the work of proving impact for visibility initiatives both unpredictable and a bit nerve-wracking.

I would venture an educated guess that a lot of the activities marketers default to for AEO campaigns can be mapped to SEO strategies. That’s not a bad thing. But search engine optimization is incomplete when applied to large language models.

Here’s what I mean.

Think about why people use LLMs instead of just typing five to ten words into a search box. It’s because they can do a full contextual brain dump — tell the machine everything about what they’re trying to figure out, with or without a clear question — and let it sense-make. The human equivalent of that is empathy. You tell someone your whole situation so they can actually understand it. People do the same thing with LLMs.

But when it comes to getting found by those same systems, we revert. We find the keywords, structure the content, build the links, copy competitors. When we track outcomes with GA4, we call it SEO. If we’re tracking outcomes for those same actions using something branded as an AI Visibility tool, we call it AEO or GEO. But the fundamental strategy remains unchanged.

What we’re not fully accounting for is that LLMs have multiple distinct behaviors. And those behaviors don’t all respond to the same strategy — because these are generative systems. The generative part doesn’t go away just because we’re using them for visibility instead of content creation. Some unpredictability is structural. What you can do is understand which behaviors you’re actually optimizing for and build a strategy that accounts for all of them.

In the absence of historical data and transparency, engineer visibility around behaviors — both your ideal customers’ and the LLMs they use to find you.

---

---

There are three LLM behaviors – ways that AI systems access knowledge – that matter for your AEO strategy.

• Recall — what the machine already knows, encoded before a conversational query starts
• Retrieval — what the machine goes out to find when it needs a current answer
• Inference — what the machine constructs when it decides it already knows enough

Those are three different behaviors. They have three different optimization requirements. And a strategy built for one of them will underperform — or fail entirely — against the other two.

Developers know these as pre-training data, RAG, and inference. That’s the architecture layer. What I’m mapping here is the founder layer — what these behaviors mean for brands and businesses that aren’t building AI, but are being evaluated, recommended, or ignored by it.

RAG and live retrieval are consolidated here because they require the same strategic response from a brand. Inference gets its own category because the visibility failure it creates is distinct, underdiagnosed, and as far as I can tell, undervalued in the marketing and brand strategy space.

Your visibility strategy has to be built around how these systems actually behave. Not around a playbook written for a different machine.

Alright – let’s get into it.

---

LLM BEHAVIOR 1: rECALL – WHAT THE MACHINE ALREADY KNOWS

Recall is what happens when a machine answers from what it already knows — no search, no retrieval, no real-time pull. This is parametric knowledge pulled from training data – the information that gets encoded within the weights and biases of a machine during training. It’s what shapes how a model understands the world, recognizes entities, and forms default answers.

Fancy technical term for real marketing consequences. Because on the marketing side of things, when a machine knows who you are without having to look you up — that’s recall doing its job.

The catch is that you don’t get to decide when that happens. Training data is updated in cycles — and the window for getting into the next one doesn’t stay open indefinitely.

This is the long game… and it’s a little bit of the wild west right now. A few days ago, I would have said prepare for that window to open once a year. Yesterday, I was thinking maybe it’s just a semi-annual cadence gap. Today, I’m thinking, “Yeah… probably annual… or something.” Just being transparent here.

I typically hedge by saying you’re trying to get into a training corpus a few cycles away — not this one, the next one, maybe the one after that. That hedge just got tighter. A 2024 Johns Hopkins study found that effective knowledge cutoffs often differ significantly from the dates AI companies report.

In some pre-training datasets, the effective cutoff aligned to 2019 despite a stated 2023 cutoff. In part, because over 80% of documents pulled from CommonCrawl were older versions, and deduplication failures allowed old near-duplicates to remain as part of the training data.

In a nutshell, we’re talking about the long game possibly being longer than I originally thought. Which makes starting now more urgent, not less.

We don’t get to choose what gets encoded in machines. AI companies don’t publish the secret sauce that makes up their sourcing criteria. What we can observe — from patterns and from what AI systems determine to be authoritative is that training data skews toward sources seen as stable and credible. Academic research. Large publications with significant subscriber bases. Research firms. Content that gets cited repeatedly, attributed clearly, across multiple sources over time.

Your play isn’t to aim directly at the training data. Your play is to show up in the kinds of places training data pulls from — consistently enough, attributed clearly enough, that when the next cycle runs, there’s a reasonable probability your ideas get encoded. OR show up alongside it

Earlier this year, I published a book that probably could have been a year-long LinkedIn newsletter series. Same content. Same thinking. I published it as a book specifically to increase the likelihood of getting it into a training corpus sooner. I built a Notion tool around the same concept for the same reason — more surfaces, more signals, better odds. That’s the training data play. You’re not controlling the outcome. You’re managing the probability.

If you’re a Marvel fan and watched (or tolerated) Agents of SHIELD, this is Sybil the Predictor territory. This is timestream logic. Chronicom science. And not nonsense this time. Sybil doesn’t control what happens. She reads probability trees and positions accordingly. You’re doing the same thing. Congratulations on your evolution.

---

LLM BEHAVIOR 2: RETRIEVAL – WHAT THE MACHINE GOES OUT TO FIND

Retrieval is the real-time layer. It’s what’s actively running when someone asks a question and the machine goes out to the web — or a connected knowledge base (think Project files) — to find a current answer. This is the layer the market is almost entirely focused on right now, and for good reason. It’s responsive. It has a fast feedback loop. Publish something today and a retrieval-first system like Perplexity or Gemini can surface it within days.

Optimizing for retrieval means being crawlable, being findable, being structured in a way machines can parse. It means your identity is understood (a canonical bio can work wonders here), your site’s pages aren’t blocked, your content isn’t buried in JavaScript, your metadata is doing signal work even when a machine decides not to read the full page.

One thing most retrieval-focused content skips: this layer is also responsible for personalized answers. That part is so important to remember.

Firecrawl’s breakdown of training data vs. retrieved data vs. live web data explains retrieval-augmented generation (RAG) as the system that gives AI agents access to specific, dynamic, or context-aware knowledge at the moment of a query.

In plain language: when someone asks an AI system to recommend something and the answer feels tailored to them specifically, that’s the retrieval layer working from what it knows about that user’s context.

Which means “best of” and “recommend me” queries aren’t just about who has the most content. They’re about who has the most relevant signal for that specific user in that specific moment. That’s the generative in Generative AI, by the way.

Full Firecrawl piece here.

If we fall back on our Marvel construct, Retrieval is Friday — Tony Stark’s AI. Always on, always scanning, always pulling what’s needed the moment it’s needed. For brands and marketers – do the retrieval work. But know it’s one system, not the whole strategy. It may be a significant part of your strategy, especially if you work in a fast-moving industry where the snapshot of the world encoded in training data hasn’t caught up to present reality, but it’s still only part of your strategy.

---

LLM BEHAVIOR 3: INFERENCE – WHAT THE MACHINE CONSTRUCTS WHEN IT DECIDES IT ALREADY KNOWS

This is the one I haven’t seen many people building for. And it’s the one most likely to get you misrepresented.

Inference, in the developer sense, just means the model generating a response — it’s a mechanical term for what the model does. But there’s a specific visibility failure that lives inside inference that has no name in the brand strategy space yet: the moment when a machine decides it probably already knows enough about you or your topic to answer without retrieving anything at all.

It doesn’t go to your page. It doesn’t throw a “fetch tool failed” error. It assembles an answer from what it’s already holding — its training data, the context of the conversation, what it can see from a distance — and delivers it with full confidence.

A few weeks ago, I built a page specifically for machines — an onboarding page so I could introduce myself to a client’s AI system. Every other system could pull it. ChatGPT wouldn’t. When I asked why, it told me it could see the page just fine. It just wasn’t parsing it. It could infer.

It inferred completely wrong. Built a detailed, confident, entirely incorrect summary. No error message. No caveat. Just wrong — and indistinguishable from right.

This is what people mean when they say AI hallucinates. But hallucination makes it sound random. It wasn’t random. The machine made a probability judgment: I probably know this. It didn’t. And that poses a problem for your visibility strategy because there’s no doubt this is happening to users who are pulling information about you, or about the thing you’re known for offering. And there’s no dashboard alert that tells you, “Hey, your dream client just pulled you up, but everything they saw was totally wrong.”

Constructing an Inference Strategy

It’s worth it to understand what’s probably happening when a machine infers. First of all, it’s not personal – so I owe the team over at OpenAI an apology for repeatedly referring to ChatGPT as obstinate. And I’m, by no means, an engineer. I’m a writer who is great at spotting patterns. And not afraid to have an entire conversation with machines about what the heck is happening at any given moment.

Inference ticks a couple of boxes at once. First, it requires fewer machine resources than retrieval. Second, it’s a risk-containment move. Meaning, the machine is filled to the brim with data that’s been cleaned, filtered, deduplicated, and scored.

Your content hasn’t.

So, if the machine comes across a query it determines it can answer without retrieving more information, it will. And if a machine is asked to go to a webpage that doesn’t convince it that there’s something on that page it doesn’t know and that it needs (in the URL, context of the website, trust/authority of the site, and schema), it simply won’t go. It will infer what the page is about based on the URL structure and what the site is about. Again – less risky than relying on its own encoded data.

The machine I’ve seen that leans the hardest into this behavior is ChatGPT.

You can’t reliably stop a machine from inferring. What you can do, however, is construct the mirror it’s likely to use when it decides not to look directly. That means making sure your most important concepts, frameworks, and capabilities exist accurately across enough external surfaces that when the machine infers, it’s working from a reflection you built intentionally, not whatever it assembled from incomplete signals.

This includes:

• Your metadata.
• Your social presence.
• Your earned media.
• Your citations.
• The corroborating signals on third-party sites that tell the machine what it’s looking at before it decides whether to look closer.

This is Uatu the Watcher — What If…? He observes everything from a distance. He doesn’t retrieve. He doesn’t intervene. He assembles a picture from what he already knows and what he can see from orbit. Sometimes the picture is right. Sometimes it’s spectacularly wrong. And he delivers both with the same certainty.

You’re not stopping the Watcher from watching. You’re making sure what he sees is accurate.

---

Build a Strategy That Covers All Three

Training data. Retrieval. Inference.

Most visibility strategies are built for one. Maybe two. The brands that hold up across platforms, across time, across the gap between what machines know now and what they’ll encode in the next training cycle are built for all three.

Training data is the probability play. Start now, show up in the right places, build toward a future cycle.

Retrieval is the visibility play. Be findable, be crawlable, be structured for machines operating in real time.

Inference is the mirror play. Seed the secondary layer so that when a machine decides it already knows — it knows the right thing.

Increase the probability that whenever machines may look, what they find out about you is accurate.

---