Algorithmic, probabilistic, deterministic… ic, ic, ic. Our recent GEO survey prompted responses from folks who mentioned terms like these as they relate to large language models (LLMs), how those models actually work and what implications these characteristics have for a generative engine optimization strategy.
Without spoiling too much, a couple of respondents had questions and doubts about how much anyone can truly influence probabilistic platforms’ outcomes.
So without further ado, let’s explore the technical details of LLMs and generative AI search platforms: how they determine what to “say” next, and how much you can really influence their responses in your brand’s favor.
How Do Large Language Models Actually Work
Several respondents to our survey wrote in noting that LLMs like ChatGPT are probabilistic platforms. That is objectively correct. But what does it mean?
For anyone who is easily befuddled by technical terms and details, the differences between these “-isctic” tongue-twister terms are much clearer to understand than they are to say five times fast:
Probabilistic relates to probability, i.e., the likelihood that something will happen. ChatGPT is probabilistic because it’s designed to choose the next most likely token (which could be a word, part of a word, punctuation mark, etc.) from a set of likely outcomes.
Based on the model, its latest training data and even how a user phrases or organizes their question, ChatGPT’s response might look different each time even though it’s answering the same question. Even the same user, asking the same question in quick succession, could elicit a slightly different response.
Forget about your pre-programmed small-talk quips, sales scripts, “Hellos!” “G-days” and “Fine, thanks. How are yous?” Those are, for all intents and purposes, pre-programmed responses that you’ve trained your idle brain to repeat whenever you walk by someone on the street (hint).
But what you say when someone asks “How are you?” versus what Sally, John or Sidney say when someone asks them how they are is most likely different.
- Me: “Good, thanks. How are you?”
- Sally: “I’m well. And you?”
- John: “Great! Thanks for asking.”
- Sidney: “Never better!”
Same question, different response, essentially the same sentiment. Humans are probabilistic. It’s likely that even you have answered the same question more than once over your lifetime using a different combination of words each time, even if only slightly.
Deterministic means a process or system produces the exact same output for any given input. They follow a set of fixed rules where input #1 equals output #1, every single time. A basic calculator is an easy example: type in 2 + 2 and, barring a malfunction, you’re getting 4.
Traditional software is full of deterministic processes — click a button and a predefined action occurs.
LLMs generate responses based on probabilities across a huge number of possible tokens. At each step, the model weighs potential continuations and generates one, then repeats the process until it has constructed a response.
And if the response is left more or less to chance, how can you control it to boost your AI visibility? Well, you can’t. But you can control your corner of the information environment that probabilistic models work from, and that is the basis of an effective generative engine optimization strategy.
Subscribe to
The Content Marketer
Get weekly insights, advice and opinions about all things digital marketing.
Thanks for subscribing! Keep an eye out for a Welcome email from us shortly. If you don’t see it come through, check your spam folder and mark the email as „not spam.“
If LLMs Are Probabilistic, Can You Actually Influence Them?
One of our survey respondents asked: “How much can we influence a probabilistic platform?”
You cannot dictate exactly what ChatGPT, Gemini, Perplexity or another generative platform will say about your company every time someone asks a question. You also can’t guarantee your brand will appear for a particular prompt, secure a guaranteed No. 1 position or lock in a specific sentence about your product. That’s just the nature of a probabilistic system.
But probabilistic doesn’t mean completely random per se, because an LLM isn’t pulling words out of a hat. The model itself, the information available to it, a user’s prompt and, depending on the platform, the information retrieved from the web at the time of the query shape the output.
Google, for example, says its generative Search experiences use retrieval-augmented generation to ground responses in relevant, up-to-date web pages. Its systems can also perform “query fan-out,” issuing multiple related searches to gather information relevant to different parts of a user’s question.
That means the exact response may vary, but the information available for the system to discover, interpret and draw from holds influence — the key term from that respondent’s question.
GEO Is About Improving Your Odds, Not Engineering an Answer
With GEO, there isn’t some secret formula you need to crack. In fact, Google explicitly recommends focusing on established SEO fundamentals, useful original content and technical accessibility rather than chasing special “AEO/GEO hacks.”
A better way to think about GEO is as an exercise in improving the information environment surrounding your brand. You can’t control which answer a probabilistic model generates, but you can control a great deal about what it has the opportunity to find:
- Clearer expertise.
- Original information.
- Detailed product and service pages.
- Curated subjects you want your brand associated with.
- A technically accessible website.
- A reputable presence beyond your own domain.
While none of those actions guarantees a particular output, collectively, they give search and generative systems more useful evidence about who you are, what you know and where you’re relevant.
So, What Can Marketers Control?
Quite a bit.
What You Know
Your business holds a lot of knowledge, no matter its size. And that’s your first advantage!
Subject matter expertise, customer insights, firsthand experience, informed opinions, and maybe even proprietary research and original data. All of that gives you something unique, not just a reworded version of what’s already online. Find those things, document them and make them accessible.
How You Communicate It
Knowing what you know, so to speak, is only phase one. Making that expertise explicit and available is phase two.
Throughout your website and pages, clearly explain what you do, who you serve and what makes your approach different. Answer the specific questions your customers ask, define important concepts, publish detailed examples and support claims with evidence. And anything that’s considered integral to your business, keep accurate and current.
LLMs might be robots, but this isn’t about writing robotic content specifically to appeal to these systems. It’s about reducing ambiguity for anyone — human users or machines — trying to understand your business.
What You Talk About
You also control the subjects on which your brand builds depth. If you want to be associated with a particular service, problem or area of expertise, one passing mention on a service page isn’t much of a strategy.
Instead, build a meaningful body of knowledge around it. That could include educational content, original research, case studies, expert commentary, videos, comparison pages, FAQs and other resources that address the topic from different angles and demonstrate genuine depth around the subjects that matter to your customers and your business.
Where You Show Up
Your own website is only one piece of your online presence. It’s a great place to start, and ironing out the kinks in your website is more or less a prerequisite since that’s most likely where users are to end up if they’re considering you.
But industry publications, professional associations, relevant directories, review platforms, partner websites and other reputable third-party sources can all contribute information about your brand.
If your website says you’re an expert in something, that’s one signal. If respected sources elsewhere on the web also discuss your company in that context, there’s a broader body of information connecting your brand to that subject.
Does that mean chasing mentions anywhere you can get them? No, because relevance and credibility matter much more than sheer volume. Modern SEO has taught us that time and again.
If you’re unsure where to start, ask: Where would you reasonably expect a trusted brand in your industry to appear? Those are the places probably worth paying attention to.
Probabilistic Doesn’t Mean GEO Is Pointless
This brings us back to another comment from our survey:
“There will not be a way to influence it, nor will there be consistent responses/answers since LLMs are probabilistic and not [deterministic].”
Indeed, generative responses won’t always be consistent. But inconsistency doesn’t necessarily make influence impossible. Marketing has rarely offered perfect control over an outcome anyway:
- SEO doesn’t guarantee that a page will hold the same ranking indefinitely.
- PR doesn’t guarantee how every reader will interpret a story.
- Advertising doesn’t guarantee that someone who sees an ad will actually buy the thing.
What marketers do is identify the factors they can affect and use them to their advantage to increase the likelihood of a desired outcome. And GEO isn’t all that different.
Rather than the goal being “make ChatGPT say X,” it should be to make your brand and expertise easy to discover, understand and substantiate when an AI system is looking for information about the subjects you care about. Then, measure whether that’s translating into greater visibility over time.
Focus Less on the Algorithm and More on Your Information
Nobody outside the companies building these models can control how they work. On top of that, how they work will continue to change with each update.
So, focus on what you can control: your expertise, what your organization knows, what you publish, how clearly you communicate it, the topics you invest in and the places where your brand builds a credible presence.
Those are durable investments whether someone finds you through a traditional Google result, an AI Overview, ChatGPT, Gemini, Perplexity or whatever comes next.
You can’t directly influence a probabilistic platform enough to guarantee its next answer … but you can potentially influence it enough to make your brand a much better candidate for these platforms to cite.
