Thoughts on the death of keyword search

(Author’s note: This is quick and dirty thinking and stream of consciousness -> text. I’m just trying to get into the blogging habit)

I wanted to title this “the death of Google search”, but it’s not dead yet, not by a long shot. What is dead however is simple keyword search, where you put in some words hoping to match something written on some webpage, and then you play around with different wordings hoping to find the information you want.

In the vein of “What do you know that everyone is wrong about?”, I have mentioned to friends for years, ever since word2vec became popular, that Google Search’s days are numbered. In retrospect, I should have gone on record making a prediction.

Here’s the deal: Google Search started sucking many years ago. People have long pointed out that attaching “reddit” to the end of a search greatly increases search quality. Google’s big edge has always been the vast amount of data it indexes. I’m not privy to the internals of Google search, but my guess is that it doesn’t have that big an advantage over Bing in terms of algorithmic intelligence. It could be a 10–20% difference (i.e. revenue would fall 10–20% if they replaced their algorithms with Bing’s), which translates to billions of dollars at large company scales, and is enough to win over customer loyalty, but it’s not enough to safeguard against disruptors.

Google’s decline started with the rise of Facebook and Twitter, as more information started to be locked away behind APIs. The “reddit” trick still works because reddit did not lock away its information the way FB and Twitter. A kind of symbiotic relationship, with Google bringing Reddit traffic and Reddit keeping the quality of Google Search reasonable. Google has the same relationship with news providers, with Amazon for shopping.

But this is a tenuous relationship. And just like FB and Twitter locked their data away, so would other companies if they had another way for their information to be found. And that would be ads on other properties that people visit, like Twitter and FB. Google’s edge is really that they provide site specific search that is better than the websites can themselves provide. All this changes with advances in ML. We could see this with word2vec, and this is what I used to point out to friends. More intelligent embeddings representing documents make any intelligence Google has baked into their algorithm moot, and the industry ripe for disruption. And we have now seen the beginnings of this with ChatGPT starting to eat Google search queries.

I have some more predictions to make:

1. Ads are not going anywhere in the near term, and are an effective way of sharing information. Humans are creatures of habit. If not the newspaper or TV or the app they open instinctively when they’re bored, there is some place they will put their attention, and ads there are the best way to reach them. Google was saved by their purchase of Android, and the deal they have with Apple using Google for their default search. This could all disappear fast.
2. Ads CAN be replaced, the biggest factor driving this will be that users hate them. We like jingles and we like the “push” nature of ads, so I can find out what new style of burger is available now, but when spending real money, we look for recommendations from friends. The replacement for ads could have been recommendations from your friends on FB and Twitter, but that did not pan out. I think what will pan out are Personal Assistants. Something that keeps track of your everyday life, and can recommend something to you on your own terms.
3. The advantages of “personalization” as it exists today disappear when the core model becomes more powerful. Any business that tries to create a “personalized profile” of you in exchange for your data to offer you better services is using it as a replacement for the low quality of content understanding that they have. One big disadvantage of current personalization approaches is the “cold start” problem, where you need to use the system for some time for it to become good.
4. On the other hand, contextual search will become more powerful. This is more like ChatGPT. Currently, you can give it a somewhat long description of what you’re looking for (the “context”), and it can list out products or services for you matching what you need. In the future, your history can serve as this context.

So, similar to what many have said in the past, I’m predicting the rise of the “Personal Shopper”: a system that knows you well and negotiates with companies selling products and services on your behalf. It can effectively “see” the ads for you and let you know if there is something that will catch your interest or if it’s something you need. It can search out things that you need as you need them, but can have a conversation with you to determine what you really need.

What I am still unclear about is the transition from our current world to one with Personal Assistants. To get the foot in the door, PAs could start out just monitoring the text you write, maybe on your laptop or helping craft the messages you send to friends. Trying to talk to ChatGPT, I have found it to be very convenient to share my thoughts via voice, so that could be a vector. But I still find it easiest to read, so my favored interface is one where I’m composing text to send via voice, being shown feedback and the final message via a display, and using touch to guide the process, e.g. pointing out where something should be said.

Time to get to work on a prototype!