How Does Generative Search Work | The Practical Marketer

Key Takeaways

So let's look a little bit at like the mechanics, what's actually going on in these generative engines. So this is a diagram from uh a GEO article, one of the earliest ones that was published and it's an academic one from university researchers on archive. Um if you have the brain for dense reading, then I recommend it. If you don't, hopefully this gives you a good overview. So this little blue guy on the left, that's us. That's the searcher. And when you put a query into chat GBT or a generative engine, it essentially formulates your query. You can put in really long paragraphs or questions and give it lots of context because then what happens is it starts to reformulate that and break it down into relevant queries. It runs those through its own knowledge and through search engines depending on the model. It comes through a summarizing model. So, it's taking all of those massive amounts of search pages and figuring out the most important information and then it creates a pattern of text to essentially give a response and output it to the user. Perhaps the most interesting thing about this model that we should be aware of though is sometimes this part in red doesn't need to happen. So sometimes these generative engines like CHBT can just answer based on their own underlying training data that they were first trained on without actually going to the web and making a search for things. Not always, but sometimes. So I've got a screenshot from Plexity because this is one of the generative engines that's best at showing you actually what's happening, like what it's breaking down. So up the top the query and this is quite a long query that I've put in as an example of the kind of things that people are more comfortable searching in generative engines. So for example, I need a mortgage and most lender options they're banks. I'm worried about the banks not being ethical and I only want to use one where my money isn't put to unethical things. For example, like the war in Israel. What the really clever thing that these engines can do is actually break down what I'm saying, what we're searching in here into different searches. So the underlying web searches that are made here, there's two of them. The first is ethical mortgage lenders and the second is banks funding the Israel war. Because what it's done is run both searches for me and then collect all of the information and it gives you the citations here. So you have a really good idea of the kind of articles and websites that Perplexity likes to learn from. And then it's taken that information to give me one answer where it's put it all together. super smart. So what does this actually mean? It means that generative engines are they're trained on different data. They have different search capabilities, but it means they also give different outputs from Google, but also from one another as well. So I've put that into this very bright sunny table to show you what's going on with three of the biggest ones. So chat GBT, Gemini, and Perplexity. So the training data, all three of them are using different training data sets. So chat GBT had the common crawl data set. Gemini, which is Google's product is of course using Google search index and perplexity uh uses third party LLMs instead uh from OpenAI and Anthropic. They also have their own crawler bots. And maybe the most interesting thing that I think is going on here is the ones that have access to the web, which all three of these do to actually make an underlying search, they're using different search engines, which is perhaps one of the biggest distinctions that we as SEOs need to make when we're optimizing. So, Chat GBT is using Microsoft's Bing because they have a partnership together. Uh Gemini is owned by Google, so it's of course using Google search and Perplexity say that they have their own proprietary tech uh and web crawling infrastructure that they're using. So they're saying they're not using either of the other searches that are going on there, which means generative engines are essentially using data sets, text patterns, and context to create us new responses every time. What they're not doing is copy and pasting. They're not pulling direct answers just from a database or copying exact words. They are using information from different sources to pull it together to create a response that is relevant to us. And increasingly, we're also seeing them start to make responses more personalized if you've got settings turned on in chat GBT, etc. to allow them to learn about you and the kinds of searches that you've