My Top Synthetic Media Prediction for 2023 May Already Be Happening
Microsoft to tap into conversational search
What do you do when your top prediction for the year may have just been confirmed in the first week of January, and you’ve been too busy publishing other things to put the annual predictions article out yet? I guess you roll with it anyway.
The Information reported yesterday:
Microsoft is preparing to launch a version of its Bing search engine that uses the artificial intelligence behind ChatGPT to answer some search queries rather than just showing a list of links, according to two people with direct knowledge of the plans. Microsoft hopes the new feature, which could launch before the end of March, will help it outflank Google, its much bigger search rival.
Why BingGPT
Microsoft has long sought to compete with Google in search. Bing’s estimated 2.95% global search market share suggests that Microsoft could benefit from a new search paradigm. Conversational search enabled by Large Language Models (LLM) may be the opening for Microsoft was looking for.
As I will describe below, conversational search has features that Google’s current search model does not offer. Google has its own LLMs, and it will likely introduce conversational search at some point. However, there appears to be a window of opportunity open today for a solution that fills in the gaps in Google search.
The market opportunity is surely a reason that would cause Microsoft to collaborate with OpenAI on this solution. Another reason to believe Microsoft would be working on this is its $1 billion investment in OpenAI. In addition, the company’s GitHub Copilot is surely the most successful application of the OpenAI Codex model for writing software code. After applying Codex to code completion, it seems natural for the companies to collaborate on search using GPT-3 or, more likely, GPT-4.
This is not to say a BingGPT combo will rival Google’s core search solution anytime soon. However, it is likely to offer benefits that Google does not and serve as a vehicle to increase Bing’s market share.
Google’s Technical Debt
Google revolutionized search when it introduced PageRank. This was scalable and relied on the wisdom of publishers, followed by the wisdom of the crowds. The measure of backlink quality was based on the assumption that publishers were incentivized to favor links to higher-quality webpages because that better served their own interests and the interests of the website visitors. This was combined with measures of user preferences for certain search results to create an automated feedback loop that enabled Google’s algorithm to become better with each search session.
This is a bit of an oversimplification, but these two ideas turned out to be far better predictors of website quality and search user satisfaction than keyword or ontological methods. But, you may be able to see the long-term weakness of such a model. It is based on the website and the webpage. PageRank is named after Larry Page, but that doesn’t change the fact that the page and site typically drive the search results page. The ten blue links have become a common way to think about Google search results.
Google has long understood this is not always the best way to serve the search user. That is a key reason why “knowledge panels,” “answer boxes,” and “people also ask” features were added to search results. What is more efficient than typing a query, receiving a list of links, choosing one, and then finding the answer to your question on a high-quality webpage? How about removing two of the steps? Type a query and receive the answer. These features have changed search.
Rand Fishkin found in 2020, 65% of all searches on Google “ended without a click,” based on a sample of five trillion searches. There is no need to click if the answer box or knowledge panel does the job. Google has been moving towards developing a more accurate understanding of page content and content quality for a long time. This typically provides a better search experience.
The best analogy is the skilled librarian. Pre-internet, a university librarian could often help you answer a question by suggesting you look up several specific books which is the equivalent to the ten blue links concept. The librarian, like Google, pointed you to a likely source where your question could be answered. Sometimes, the librarian simply knew the answer and could pull a single book for you and show you the exact passage with the information. In one case, they facilitated your research. In the other instance, they essentially did the research for you.
Some librarians, however, had another skill that Google still lacks. They could sometimes answer your question by pulling from two or more sources.
Integrated Search
This is where integrated search comes into play. Google can generate an answer box if the information exists on a single page somewhere on the internet. Knowledge panels can source information from multiple sources but are limited to people, places, organizations, and things. These are both about facts or what Google’s knowledge graph deems to be facts. Otherwise, Google can only provide you with information related to your question that has already been published somewhere.
ChatGPT has many redeeming qualities. However, one thing that intrigued many people was its ability to synthesize information, seemingly from multiple sources, to answer questions that may not have previously been addressed on a single webpage. This was a reason why I would sometimes use InstructGPT earlier in 2022 for searches where I was unsuccessful with Google. It is the same reason I am testing YouChat and other conversational search engines now.
Large language models are not limited to the vessel (i.e., page) construct. They are about the words and, through a prediction model, the ideas they are used to express when assembled together. This is more humanlike. We can easily integrate data from multiple sources to answer a question or generate a new idea.
Conversational Search
Another feature that caught ChatGPT users’ attention was the conversational nature of the search. You can enter a search query in natural language for Google, but there is no ability to have a conversation about the answer or refine the search through collaboration. ChatGPT and, to some extent, YouChat, can continue expanding on the ideas expressed in the question and subsequent answers. This can lead to the discovery of new information and also can help the user refine their search if the answers are not quite what they wanted.
This is how we would interact with the librarian mentioned above. We could collaboratively refine the ideas in the question to arrive at a better answer. Isn’t that the entire point?
The conversational search options we have seen to date are generally brittle and lacking important features. With that said, they are also a powerful addition to the traditional search model pioneered by Google because they provide new benefits. The model is very similar to what we would expect from having an expert sitting beside us.
Based on what we have seen so far and the widespread consumer interest, it seems inevitable that some types of search will be conducted in a AI-enabled conversational format. Microsoft seems well-suited to be a leader in this space.
