How LLMs actually work (the non-technical version)

I'm not going to explain attention mechanisms or token embeddings. You don't need that. But there are three things about how LLMs work that directly affect whether your brand gets mentioned.

When someone says "ChatGPT knows about your brand," that's not quite right. ChatGPT was trained on a massive corpus of text — books, articles, websites, forums, code, and more. During training, it learned patterns about what text looks like and what words tend to appear together.

Your website isn't in ChatGPT's memory. The text from your website — or more likely, text from other sites that mention your website — is what the model encoded during training.

This is why getting mentioned on other sites matters more than optimizing your own site. The model doesn't visit your page. It learned about you from what others wrote.

When ChatGPT answers "what's the best email marketing tool," it's not looking up a list. It's generating text one word at a time, choosing each word based on what's most likely to come next given the context.

The model has learned that when people ask about "best email marketing tool," certain brand names tend to appear in the answer. It's pattern matching, not reasoning. If your brand name appears frequently enough in the training data for that context, the model will generate it.

This is why frequency of mention matters so much. The more often your brand appears in relevant contexts, the more likely the model is to include it in generated text.

Every model has a training data cutoff date. GPT-4 was trained on data up to early 2024. Claude has its own cutoff. Gemini has another.

This means the model's knowledge is frozen at a point in time. If your competitor launched a major product three months ago, the model doesn't know about it. If you had a big PR push last month, the model doesn't know about that either.

This creates interesting dynamics. Brands that were heavily discussed before the cutoff have an advantage. But it also means there are windows where being current in the real world doesn't translate to AI visibility — and vice versa.

These three facts lead to practical conclusions:

Get mentioned in training data sources. Review sites, forums, industry publications. These are the sources that feed the model.

Be consistent over time. The model encoded patterns from months of data. One article won't move the needle. Consistent presence does.

Don't expect instant results. Changes you make today won't affect ChatGPT until the next training cycle. The mentions you earn now will pay off in future model versions.

Understanding these basics is enough to build an effective GEO strategy. You don't need to understand the technology. You need to understand the incentives — and optimize for them.