How Would You Sell Ads on an LLM?
Researcher in Machine Learning, Causality and Recommender Systems. On a career break to write a book on Reward Optimizing Recommender Systems. Previously led a team of researchers and engineers at Criteo.
It seems that there is widespread agreement that there may be an AI bubble (read large language model / LLM bubble). I am not sure what to think of Ed Zitron’s commentary about the financial viability of the industry. Mariia Soldatenko makes similar observations but more specific to LLMs competing with search for advertising revenue.
While the singularity is probably not happening next week, and we are not all going to be made redundant or attacked by SkyNet, it does seem like a lot of people are using LLMs as a replacement for search, and for now this is paid for by venture capitalists. Could LLMs follow search by making money from ads (despite their much higher running costs)?
Firstly, how does Google make money by advertising on search? This (rather old) video from Google explains it. While on paper Google bills users for clicks, it uses a somewhat complicated mechanism to effectively determine the auction winner as the highest cost per impression. Broadly, the auction winner is determined by determining who has the highest bid multiplied by the click through rate (Quality Score in Google parlance). This effectively means an advertiser can win the auction by being both relevant (with a high click through rate) and having a high bid. A low bid might be enough to win the auction if the ad is very relevant, if an advertiser is willing to pay enough it can show low click through rate ads (presumably low relevance ads). A second price mechanism encourages advertisers to bid their maximum price, as they only pay the second price of the competitor they beat.
All of this is done at a keyword level. That is bids and ads are tailored to particular words like “insurance” or “coffee near me”. Some of these keywords (like “home insurance”) are actually very sought after and this causes a market to operate on these keywords. From the perspective of Google, there might be (say) one thousand searches of “car insurance” on a day, and it will seek to sell the ads on these searches at the highest price.
This is (one) obvious place where LLMs are very different to search engines. A user is likely typing in much more complicated text into the LLM than just “home insurance”, the user may be asking a question about the types of cover in their area, the price of competitors, the session will be long and may contain multiple queries asking for refinement. The complexity and uniqueness of the session makes it much more complicated to create a market where different advertisers bid against each other. The additional information however might help determine which advertiser would be most interested in this opportunity, and what type of ad it should show. IN short, hyper-personalization brings with it both opportunities and challenges.
It may be possible to embed the text and provide this embedding to the advertisers’ bidding engines, yet such a formulation makes it difficult to combine relevance signals with a click model. Another alternative could be that the LLM operator could simply segment the users and run auctions on the segments. Given the simplicity of this solution, this seems quite likely, but the advertiser would not know the nature of the users’ conversation (say) about insurance, only that it was about insurance in some sense. This problem can surely be overcome, but I don’t think we know what it looks like, assuming that LLMs are (profitably) able to replace search for some use cases.
