but what is the impact and possible countermeasures?
Brief History, Google’s Use of AI and Current Situation
Google have an incredible AI competency, in particularly since they bought DeepMind in 2014. It has been predominantly oriented towards improving internal capabilities. Last year (2022) Google demonstrated that AI improved data center energy efficiency by 40%. They have used AI for ad optimization and ad copy optimization. The YouTube follow video algorithm have been improved by AI, so that a highly relevant video is suggested to you, as your next video to watch, which massively increased YouTube hours watched per user. This in turn massively increased YouTube revenue.
Google has also leveraged AI to determine the right time and place to insert ads in videos on YouTube, AutoComplete in Gmail and Google Docs. Thus far not much of this competency have been oriented specifically to avoid disruption in search. Obviously with ChatGPT entering the scene as an alternative to ordinary search, things have come to a crossroad.
One can think about traditional internet search as information retrieval. You pull data from a static data set, and it involves scanning/crawling that data set, creating an index against it and then a ranking model. Based on what you are querying, data is pulled out of the index to present the results from the data that is available. All of that is typically done in a 10th of a second. If you think about the information retrieval problem, you type in the data or some rough estimation of the data you want to pull out, and then a list is presented to you.
Over time Google realized that they could show that data in smarter and quicker ways. They could identify that you are looking for a very specific answer, and reveal that answer in the “one box“, which sits above the search results. If you asked “When does this movie show at this theater?“, they can pull out the structured data and give you a very specific answer, rather than presenting you with a list. Over time it turned out that other kinds of modalities for displaying data were more beneficial than lists as well, such as maps, or a matrix result for shopping options. The way that information retrieval results were presented to you improved to be more fit for purpose and it created a much better user experience.
The challenge with what this new modality that ChatGPT represents, is that it is not really fully encompassing. If you consider the human computer interaction challenge, you want to see flight times, airlines and the price of flights in a matrix, rather than a text stream written to you. The same goes for shopping results. If you are looking for various people’s commentary because you are looking for different points of view on a topic, rather than just getting one answer. However, there are certainly a bunch of use cases for which the ChatGPT type of answer (natural language) would be the preferred mode.
Considering Cost of Search and Its Future
Revenue per search done on Google typically range from 1-10 cents, and we know that roughly 1 out of 100 people click an ad, and that is where the money comes from. If we assume 5 cents, then roughly 50% margin on that search, which means a 50% COGS (Cost Of Goods Sold)/cost to run that search and present those ads. Based on that, a Google search costs them about 2.5 cents. A recent estimate on running the GPT-3 model for ChatGPT is that each result costs about 30 cents of compute. Thus, it is about an order of magnitude (10x) higher cost to run a ChatGPT search as compared to a Google search. This means that the cost of running ChatGPT like searches needs to come down by about an order of magnitude.
However, there are a lot of thoughts on how to get a 10x reduction in cost on running these models. Optimization in software and models, how you run them on a compute platform, the type hardware used for the computation (which chips etc.), etc. Currently, quite a lot of work remains before this search modality becomes truly economically competitive with Google. When you get to the scale of Google we are talking about spending 8-20 billion USD per quarter just to run search results and display them. Thus, for ChatGPT type solutions on Bing, or elsewhere, to scale and to use that as the modality across all search queries (which likely will not make sense, ref. the above paragraph), we are talking about something that today would cost minimum 80 billion dollars per quarter to run from a compute perspective, which is astronomical.
If we look ahead, the ability to optimize ChatGPT-like search queries is likely to dramatically improve in the coming years. Also, the cost of energy is reducing rapidly. When considering these to factors combined, it is very likely that this kind of search modality will become economically viable, even at scale, in the coming years.
Counter Measures
ChatGPT is an incredibly important step forward, however, it is merely a building block of in the search platform that is very likely to rapidly get commoditized. All the big companies that has a stake in search will compete over time, and is likely to utilize this kind of search modality in their search services, and other relevant services. Microsoft is currently taking the next natural step for a company on the outside of search to reach for Google’s ~93% share of the search category. Microsoft effectively bought almost 50% of this building block/tool for 10 billion USD (through their acquisition of shares in OpenAI). Microsoft aims to make this tool as pervasive as possible across their application portfolio, to render consumer expectations such that Google is forced to compromise the quality of their business model, at least in the short-term, in order to compete.
Currently, Google’s answer to this is to rush the release of Bard, a service that seems to be far from ready for the market. This resulted in a share price reduction of 500 basis points, leading to a market cap reduction of 100 billion USD. It is also not unlikely that Meta would like to take a position in this new search modality, which would further reduce Google’s share of that market. Tencent could do the same.
AdSense is typically paying out 70 cents on every dollar to the publisher. It is pretty generous, and it is the way they have kept the competitive moat. Google bid on everything and they always win, because they always share the most revenue back. As a result, they own that market with respect to acquiring content.
What would the appropriate counter measure for Google be to preserve their moat? Google could e.g. double the TAC (Traffic Acquisition Cost: payments made to affiliates and online firms that direct consumer and business traffic to their websites). It would effectively be a way for Google to guarantee an exclusivity on search traffic. This is not as far-fetched as it might sound, considering e.g. that Google pays Apple ~18-20 billion USD for having exclusivity to search for all Apple devices (and preventing Apple from developing their own search service). Raising TAC in order to preserve their moat would probably be a better alternative than to loose a significant amount of market share, say 5-20% in the coming years (in case Bard does not take off, and Chat-GPT powered Bing as well as other players starts to capture search market share from Google). If Google were to pay publishers two times more than what anybody else is paying them, this might also put them in a position where they could prevent AI agents crawling a range of websites (e.g. through a txt-file that prohibits AI agents from crawling the site, similar to the Robots.txt for search to today).
Potential Changes in Monetization Models
We may have to consider that the whole monetization model needs to change. In this new modality of search you are not looking for a list of 10 – 20 links, you are just looking for the answer. So, where is the opportunity to advertise against that? Maybe you can charge something similar to an affiliate commission if the answer contains a link or similar? But then you have to ask yourself, does that distort the best answer? Am I really getting the best answer, or am I getting the answer that someone’s willing to pay for?
Google has had a very finely tuned balance between presenting the first two or three paid ads, above the actual search results. These paid links links might actually give you a better answer than the content below them, but if ChatGPT tells you this is the top three e.g. televisions, hotels, etc. you do not need to click. The ad click model is gone.
At Google a key metric has typically been the “bounce back rate“. This is when a user clicks on a result on the search results page, what percentage of the time they come back to search again. This speaks to the quality of the result that they were given, because if they do not come back, it means they ended up getting what they were looking for. If someone created an ad, paid for it and the user clicked on it and came back with less frequency than if they clicked on an organic result, that meant that the ad quality was higher than organic search quality. This lead to the ad being promoted to sit at the top of the search result. It became a really important part of the equation for Google’s business model. That in turn would be how Google monetized more search results where they could get advertisers to pay for a better result, than what organic search might otherwise show. Thus, it is actually better for the user in this case, than say just getting an answer from an search powered by a Large Language Model (LLM).
Many search queries are commerce intention related (e.g. buying a television, gaming console, flight). That series of queries may have a very different kind of modality in terms of what is the right interface versus the ChatGPT interface. Ideally you would want to search solution to understand your intent with the query and select the relevant modality to present the result in, which e.g. in some cases would mean invoking a ChatGPT like mechanism (LLM) and in others display a ranked list, a map, a matrix, carousel or something else that. Thus, Google could do a simple analytical exercise in terms cost and user intent (better user experience: best result for the user), and that is ultimately what will kind of resolve to the better business model.
Other Promising Developments and Closing Words
A lot of amazing AI tools keep cropping up. One of them is Galileo AI which creates delightful, editable UI designs from a simple text description, which empowers you to design faster than ever. E.g. you type in “a way for people to change their name phone number and password“, that classic screen on any app, and immediately it presents you with sleek design. Another example is Midjourney which offers a bot that can generate all kinds of unique and stunning AI art and even photography quality images based on text prompts, a pre-selected image or a blend of pre-selected image and the combination of all of these.
Yet another example is the GitHub co-pilot (reminder: GitHub was bought by Microsoft in 2018) which as you are writing your code, fills in your code. It knows what you are writing and autocompletes it, just like in an email. This may enable people to e.g. build the MVP for their startup by typing in text and then publishing it. You may not even need a developer initially for your tech startup. That is truly transformative!
At this stage most LLMs that run on the same corpus of data is going to resolve to the pretty much the same answers. As more companies builds their own LLMs the advantage quickly gets competed away, unless you have access to a certain data set that others do not. Such as Microsoft having data from LinkedIn, Meta have access to the data from Facebook, or a company got the exclusive rights to Quora, X, etc.
In any case, the deployment and widespread utilization of LLM models are going to lead to economic productivity. Companies using these tools are going to reduce their costs of running their businesses, and their total net profits go up. This is what happens with every technology cycle. It always yields greater economic productivity and that is why technology is so important to drive economic growth.
Leave a Reply