Molly Ploe

SEO has always had a mythology problem. Ask ten different marketers what actually moves the needle on rankings and you’ll get ten different answers, half of them based on vibes rather than evidence. Now that generative AI has crashed the party, it’s doing the exact same thing to our industry, just faster and with way more confidence.

So our CTO, Stewart Snow, hosted a Lunch & Learn to set the record straight: what LLMs actually are, how they work under the hood, and what that means for GEO and SEO. Stewart’s got the receipts here: nearly three years hands-on with OpenAI and Anthropic tech, plus 25-plus years in SEO going back to when the term barely existed.

Here’s everything we took away from his session.

1. Isn’t an AI Model Just the Same Thing as ChatGPT?

Nope. This one trips people up constantly.

ChatGPT and Claude are apps. The LLM — the model — is the brain running underneath the app. There isn’t just one model, either. OpenAI alone has dozens, and they all behave a little differently depending on what they were built for.

Knowing the difference matters, because a lot of what people assume “AI” can or can’t do actually depends on which model is doing the work. Overestimating what the system can do on its own risks underestimating the importance of fresh data or careful prompting.

2. Do AI Models Know What’s Happening Right Now?

This is probably the single biggest myth Stewart wanted to bust, and he proved it live: he asked an older GPT model who the U.S. president is. Its answer? Joe Biden. Very confidently.

Why? Because building a model means feeding it a giant scrape of the internet and then running months of processing on it. As Stewart put it, once that process wraps, “these things are fixed. They are static. They are one-offs.” Whatever happened after its knowledge cutoff date, the model simply doesn’t know, and it won’t tell you that, either.

That cutoff has real consequences for you. Launch a new product today and an existing model won’t recognize it tomorrow. The model can discuss what it knew during training, but it needs outside help — often a live web search tool — to speak intelligently about anything that appeared after that cutoff.

3. Will AI Give Me the Same Answer Every Time?

Also no, and this one’s actually intentional.

At their core, LLMs are next-word guessers, pulling from a range of statistically likely options rather than locking onto one “correct” answer. Ask the same question five times, get five different answers. Stewart demonstrated this by asking a model to name the best content marketing agency in the UK. Guess what: The answer changed nearly every single run.

If you’re coming from the SEO world, this is a mental shift. Google gives you consistency. AI, by design, doesn’t.

4. Why Does AI Sometimes “Think” Before Answering?

To overcome fixed knowledge and randomness, AI systems add two powerful layers:

  • Reasoning: A chain-of-thought process that forces the model to consider steps before it answers.
  • Tools: Functions such as web search that fetch fresh data during your session.

You’ve probably seen that little “thinking” indicator pop up in ChatGPT or Claude. That’s not just for show.

Reasoning models are built to work through a problem step by step instead of blurting out the first plausible answer. It’s basically the AI equivalent of showing your work on a math test.

Stewart showed the difference directly: older, non-reasoning models flub simple stuff (how many R’s in “strawberry”? A shocking number of models get this wrong). Reasoning models, forced to slow down and think it through, get it right far more consistently.

Marketers should keep this aspect in mind; reviewing the “thought process” that led to an output helps you think critically about the information the LLM gave you.

5. If AI Doesn’t “Know” Current Events, How Does It Answer Questions About Them?

Tools. Specifically, web search.

Since the model itself is frozen in time, AI apps lean on tools to go grab live information. Stewart went back to that same outdated model, gave it web search access, and asked again who the president is. This time, it searched, found the right answer, and corrected itself, all while its internal “brain” still thought it was 2023.

The model still knows nothing new. That difference matters more than you’d think. Whenever you request “latest benchmarks” or “today’s news,” you’re really querying the search layer. If the answer lacks citations, assume it came from stale memory and double-check it.

Guiding the search term yourself — for example, adding “2026 study” or “site:.gov” — can tighten accuracy. These tool dynamics lead directly into visibility. If AI search functions can’t see your pages properly, they can’t retrieve your expertise for users.

6. Is GEO a Totally New Discipline?

Here’s the big one, and honestly, it’s good news. Stewart’s own takeaway: “A huge portion of GEO is really SEO ranking.”

Why? Because AI still has to search somewhere. ChatGPT has leaned on Bing. Gemini uses Google. Anthropic has used Brave. Increasingly, it looks like AI providers are quietly building their own proprietary search backends, which means we’re all optimizing for an environment we can’t fully see.

That isn’t to say that models actually reference SERP positions. They keep citing your page only when their search component repeatedly finds it for related prompts.

So, if you’re already ranking well across the major search engines, you’re probably doing fine in AI-generated answers too, because none of these AI companies have Google’s 25-plus years of search infrastructure. They’re mimicking proven playbooks, not reinventing the wheel.

That also means that if your crawlability, link equity or on-page clarity decreases, you’ll feel the impact in both classic and generative search.

7. Does My Metadata and Markup Help AI Understand My Page?

This was maybe the most actionable point in the whole session: AI doesn’t see your page the way Google’s crawlers (or a human) do. No title tags. No meta descriptions. No schema. No JavaScript-rendered content at all.

What AI actually reads is closer to a plain-text, Markdown version of your page: headings, body copy, links. That’s it.

Stewart showed us a two-step trick for checking this yourself: 

  1. Copy your page’s text into a tool like pastetomarkdown.com to strip out the things AI can’t see.
  2. Use a tool like markdownlivepreview.com to view it from an LLM’s perspective.

If your important content depends on JavaScript to load, there’s a real chance AI has never seen it. It’s the same blind spot SEO dealt with a decade ago.

Imagery also poses a potential problem for AI systems. Though some models can “see” and “read” imagery, it’s unlikely this is happening on your page during a search.

“I’m reasonably sure AI doesn’t actually bother reading these images because it’s actually extremely costly and intensive for it to do that when dealing with web pages,” Stewart explained.

8. Can I Trust the “AI Visibility” Numbers These Tools Are Showing Me?

Take them with a healthy pinch of salt.

AI providers don’t publish data on what people ask their models (understandably, considering some of those conversations get pretty personal). So most third-party tools claiming to measure “how often you show up in AI answers” are estimating, often by asking AI what search terms it’d use for a topic, then correlating those to existing Google search volume. It’s a reasonable proxy. It is not real usage data.

9. Why Does AI Know My Brand but Not My Competitor’s?

Comes down to how much of the internet talks about you and how much a model is willing to admit it doesn’t know. As Stewart pointed out, “AI is desperate to please. It really, really wants to answer your question,” rather than admit it’s coming up short.

Newer brands get guessed at, sometimes right, often not, simply because there isn’t enough reference material out there (like mentions, backlinks or coverage) for a model to draw from confidently. Brands with a long, well-documented history online (Brafton’s been around since 2008) show up far more reliably, because there’s simply more signal baked into the training data.

GEO Isn’t a Mystery, It’s a Sequel

If there’s one thing to walk away with from Stewart’s session, it’s this: GEO isn’t some brand-new discipline you need to learn from scratch. It’s SEO’s next chapter, with a few AI-specific wrinkles layered on top. Write for text-based readability, structure for quick retrieval, and keep building the authority and relevance that’s driven rankings for years.

The tools are evolving. The fundamentals really aren’t.

Thanks to Stewart for taking the time to demystify what’s actually happening under AI’s hood — and for confirming, once again, that when you ask AI who the best content marketing agency is, it still knows to say Brafton. Most of the time, anyway.