Skip to content

Generative AI vs. Traditional Search: Technical Differences

By Matthew Edgar · Last Updated: April 02, 2024

Increasingly, Google and Bing are including responses from generative AI tools in search results. These inclusions are part of a broader transition from traditional search (search as we know it today based on a list of websites) to conversational search or chat search (where generative AI responds to search queries as part of a conversation).

As we think about this transition, it is important to understand how both search engines differ from generative AI at a technical level. What makes a search engine different than a generative AI system? Are Google and ChatGPT doing the same things when they respond to a query or prompt? Or are they different and, if so, how?

The short answer is there are differences between the two technologies. By understanding the fundamental, technical differences between traditional search and generative AI, we can better understand why conversational search is so different than the current way Google operates. Understanding this also highlights how we may need to adapt our SEO strategies.

In this article, I’ll review how search engines and generative AI work at a high level. I’ll also discuss some of the ways this begins to change how we think about SEO and how we may be able to optimize for conversational search.

How Search Engines Work (A Simplified View)

There are three main steps of how search engines operate: crawl the web to find information, process and evaluate that information, and then return relevant results in response to search queries. Let’s review how Google performs these operations to understand these steps in more detail. The overall process is summarized in the graphic below from Google’s help documents.

Googlebot flow chart – source

The first step of the process is crawling the web to find information. They crawl websites with a robot; Google crawls using several robots, the primary of which is called Googlebot. Crawling begins with Googlebot discovering a new URL. URLs can be discovered from many sources, but internal and external links continue to be the most common discovery sources. Once discovered, the URL is added to a crawl queue. At some point, the page bubbles up to the top of that queue and Googlebot fetches the HTML code from that URL. Once fetched, Googlebot saves the information.

Next, Google begins processing the fetched HTML code. Google starts this process by separating the different parts of the page’s content and resolving any errors encountered within the HTML code. After this initial processing, Googlebot adds the page to a render queue. Rendering involves processing CSS code that defines how the page is styled and running any JavaScript code used on the page. The JavaScript code might add text, images, links, and more to the page. After rendering, Google finally has a complete look at the page.

After processing and rendering are complete, the rendered HTML is then passed over to Google’s indexing systems. The indexing systems begin analyzing the page’s content. This is where Google assesses what topics the page discusses, whether the content meets quality standards, determines what type of search intent this page could satisfy, and more. Based on this evaluation, the page is added to Google’s index. Many pages are rejected and cannot be indexed. Through all this processing, Google has created an index that is a sophisticated representation of all the information gathered from the crawl.

Once the page is indexed, the page is eligible to appear in search results. When a user conducts a query on Google, Google searches the index to find which pages are most relevant to the query. Google’s algorithms then decide in what order to rank the relevant websites. For some pages, Google’s algorithms may choose to extract some content from the page, such as some snippet of the page text or an image from the page itself. The results presented are referred to as organic search results and these results are based on Google’s algorithms evaluation of all known websites. Along with organic search results, Google will also present ads that companies can pay for; ads are evaluated in significantly different ways than organic search results.

While Google’s algorithms rely heavily on AI, this is different from generative AI. The algorithms are not attempting to create new content. Instead, search engines are designed to return a list of what already exists. That list of websites needs to be related to the user’s search query—the more related it is, the better it is. AI is used to ensure results are highly related. For example, Google began using an AI-based algorithm called RankBrain in 2015 that helps them better understand “how words in a search relate to real-world concepts.” Google also uses language models, including Google’s BERT (Bidirectional Encoder Representations from Transformers) and MUM (Multitask Unified Model), to help Google’s algorithms understand context and meaning. AI is also used during the evaluation of websites to determine what pages should or should not be included in the index.

How Generative AI Works (A Simplified View)

Generative AI uses a Large Language Model (LLM) to formulate responses. An LLM is a sophisticated model that represents human language. OpenAI uses the GPT language model, Google’s language model is PaLM 2, and Anthropic has a family of models referred to as Claude 3.

The first step of the process is to train the LLM on how words are used in a language. To do this, the LLM must be trained using a large collection of documents. Languages are complicated so the LLM must be fed as many documents as possible to understand all the nuances of how words can be used. For a general-purpose LLM, like the ones used by OpenAI or Google, also needs to be trained on documents spanning as many different contexts as possible because language is used in different ways depending on the topic being discussed.

The crucial thing to remember is that the LLM is not trying to understand how to organize the information contained in the training documents. Unlike Google, an LLM is not trying to retain key facts or figures. It is not a database.

Instead, the LLM is trying to understand how words are used together. It detects the patterns within the language and then builds a model that represents those patterns. These patterns are used to make predictions about how words are connected to each other; the more successful patterns result in more accurate predictions.

To make it easier to detect and process these patterns, the LLM does not store information as words or sentences. Instead, it creates embeddings. Embeddings represent words, or parts of words, as numerical values. Working with numbers allows for easier computation by a computer. The embeddings capture semantic relationships and similarities between words, so words with related meanings will have embedding values that are closer together numerically.

The model being trained is a neural network. The neural network is a collection of millions of processing nodes, which are similar to neurons in a human brain. Each of these nodes is connected to many others, and they pass data back and forth between them. As the model learns from the training data, it adjusts the strength of those connections through numbers called weights. The weights determine how much influence each connection has over the others.

The patterns being learned are distributed across this whole interconnected network of nodes and their connection strengths or weights. During training, the neural network learns how to adjust these weight parameters to better model the language so that it can make better predictions about using that language.

When a prompt is entered, the LLM creates an embedding to represent the prompt. That is, it represents the prompt as a numerical value. It then performs operations on the embedding. The prompt is then sent through the trained neural network. The neural network uses all the weights it has calculated during training to make a prediction about which word should be generated next.

The LLM does not return a single prediction. Instead, it returns a ranked list of predictions. In a great article explaining how ChatGPT works, Steven Wolfram uses the following example of how the LLM makes predictions. There is a 4.5% chance the next word generated should be “learn” and a 3.5% chance the next word generated should be “predict”.

How LLMs make predictions – source

It does not pick the highest probability word each time. To make the response sound more natural and creative, the generative AI system will sometimes pick lower-ranked words. The amount of randomness in deciding which word to use is defined by the temperature parameter. The temperature is on a scale of 0 to 1, where 1 is the most creative. ChatGPT’s default temperature is 0.7. More randomness or creativity allows ChatGPT to make more interesting predictions.

After generating the first word of the response, the AI system goes back through the neural network to generate the next word and then the next word and then the next word and then the next word, and so on until a full response has been generated. (To be precise, it is generating tokens, which can be a part of a word.)

Optimizing for Generative AI Search

At the most fundamental level, a search engine’s goal is to return a list of relevant pages related to a query. Search engines invest a lot of time and effort into evaluating websites to make sure they return the best list of results possible.

In contrast, at the most fundamental level, generative AI’s goal is to generate new content based on the patterns and knowledge it has learned from a vast dataset. It uses a large language model (LLM) to predict which words should be included.

Yann LeCun, a deep learning expert, was quoted as saying LLMs do not have “some sort of objective other than just satisfying statistical consistency with the prompt.” That underscores the differences between traditional search engines and generative AI: generation of new content based on a language model versus the retrieval of existing content from a sophisticated knowledge base.

How search engines compare to generative AI chatbots
How search engines compare to generative AI chatbots

Having a conversation with ChatGPT and asking it for information about a subject is not really the same thing as entering a query on Google. ChatGPT and the other systems are not trying to find accurate or relevant information like Google does when it searches the index and returns the traditional search result rankings. Instead, generative AI is simply making predictions about how words are connected to each other. The fact that the generated words happen to contain accurate information related to the prompt speaks to how well the LLMs are trained.

Understanding the differences changes how we should think about optimizing for a new world of conversational-based search compared to how we optimize for traditional search.

Optimizing for traditional search requires showing Google that your content exists and is relevant to a particular search query. You also want to demonstrate that your content is trustworthy and authoritative. This helps Google construct an accurate knowledge base.

Optimizing for generative AI is different. The LLM is trained on billions of records and is looking for patterns across those. To optimize for an LLM, you would need to update a large majority of those billions of records so that the LLM is more likely to make predictions based on information contained on your website.

Simply put, optimizing your website is simply not enough to make a difference in how the LLM makes its predictions. Instead, to optimize for an LLM, you need to optimize the world. That requires far more than SEO, including PR, branding, and reputation building.

What To Do Now

Traditional search engines and generative AI are fundamentally different technologies. Search engines are purposefully designed to help us find information. Generative AI happens to help people find information, but its main goal is to create new, unique responses related to a prompt.

The differences between the technologies change how we need to optimize our websites. How we optimize our websites for LLMs is not clear yet, unlike SEO where there are clearer optimization steps to follow. What is clear is we need to ensure that our company and website are widely discussed so that the LLMs can learn to model words about our companies appropriately during the training process.

As a place to start, you need to find out what the LLMs currently understand about your website or business and if LLMs are currently making predictions involving your business. Go to ChatGPT, Gemini, Claude, Copilot, and other tools and ask those tools questions about your company, your products, your services, your key staff members, and so on. Ask these tools about your competitors. What information is being returned by the LLMs? Is it accurate or not? If it is not accurate, ask yourself why. What did the LLM find in its training data that led it to make an inaccurate prediction?

Need Help?

If you need help analyzing LLMs and understanding how it is making predictions about your business, please contact me.

You may also like

What Are Generative AI Hallucinations?

One of the biggest problems with generative AI responses is hallucinations. What are hallucinations? Why do they happen? More importantly, as generative AI starts to dominate search results, how do we factor hallucinations into our future SEO strategies?

Will People Click Links in AI Overlays?

Will people click links in Google’s AI overlays? How many clicks could we lose if generative AI responses replace traditional search results?

Understanding Prompt Types

How do we optimize for prompts and make sure our companies show up in generative AI search responses? How is this different than optimizing for search queries? It begins by understanding what types of prompts people use.