We need to talk about AI search.
We care about search.
At its heart Open Opportunities is a search engine. So we watch the market, explore the technology and examine the potential of new technologies in search. The latest trend in search engines is GEO (Generative Engine Optimisation) a derivation of SEO (Search Engine Optimisation) its proponents argue that they can help your company get found by LLMs (Large Language Models) such as ChatGPT.
Getting your product referenced by ChatGPT or Gemini happens in one of two ways. The first way is because information about your product was embedded in the model, the second is because you showed up in the model’s search engine results.
What embedding means for AI search
Being embedded in the model means that somewhere in the enormity of the statistical data in the LLM your product is mentioned. If your product is an iPhone that’s a certainty, for less well known products like Open Opportunities you’ll only be in the model if your product is mentioned in the training data used to create the model. (The bigger the model the more likely you are to be mentioned).
Understanding LLM search
If you didn’t make it into the model, you’ve still got a shout of being found because you might make it into the search results that the model requested.
Once the source data and the search results have been combined, the LLM will evaluate the results to look for meaning and matches. So having a page that says “best boutique hotel Copenhagen” doesn’t mean that the LLM will perceive that to be statistically relevant. It certainly makes it more likely but this is only one of the sources that will be considered by the LLM.
Diverse sources
Search engines are one dimensional, they use one input and deliver a list of outputs. LLMs combine all their sources and then generate varied (and inconsistent outputs). LLMs use their trained data; web search; the user’s context and often the user’s chat history.
All of these factors will weigh in on the LLM’s assessment: the context of your chat, any stored data that the LLM knows about you, whether different sources also refer to your product in positive terms and whether or not rival hotels have more statistical density in the model.
So, if the user is travelling for a conference the LLM will look for a hotel near the conference venue and of the web is full of stories about bed bugs in your hotel, the LLM will consider that too.
What does this all mean?
Basically that everyone selling GEO tracking services for LLMs is having you on. Literally every single one of them. Here’s why.
1. No one can tell how the LLM ‘found’ you. Was it web search or training data? If they don’t know what the LLM did they can’t really intervene to boost your company.
2. You can’t be inserted into the model, you have to hope that the next round of training considers your product to be relevant.
3. Relying on an LLM’s search engine results is… <drumroll>… just SEO.
4. Because a user’s context is unknown there’s no reliable mechanism for ranking an LLM’s desire to mention your product.
5. The more an LLM knows about a user the more refined the answers will be and the harder it will be to predict the LLM’s response.
The answer
Have lots of genuinely interested people say genuinely nice things about your company.
Who knew?
