In the next two to three years LLMs will transform to allow advertisers a direct line into the models output. This will again upend the current winners and losers creating risk and opportunity.
The mantra of Silicon Valley the past thirty years has been to grow at any cost, monetize later. There’s some debate these days if that will change, but for the current crop of AI companies it holds true. OpenAI, Anthropic, Perplexity, etc. are all focused on getting there fastest, with “there” being a combination of AI capability and market share. Profitability comes later.
Let’s talk about the later part for a moment. Currently, many AI companies are primarily using a charge per use (per API call or per number of tokens used) or a flat rate (e.g. monthly fee) model. These companies have also invested billions in their R&D which they need to capitalize.
Pioneers aren’t always the winners. For example, VisiCalc invented the spreadsheet, but Excel won that race handily. We saw the first human genome sequence cost approximately $2.7B. Today it costs less than $1,000. The first commercially viable LLMs cost billions but their cost is dropping rapidly. What does this mean for monetization? Subscription models will be challenging under commodity systems. Enter advertising.
Advertising feels like the lowest common denominator that all tech inevitably sinks to. (And by tech I’m not just talking about software.) The search engines, the closest technology from a behavioral standpoint to today's LLMs in that people use it to find answers, quickly realized they needed to not charge users directly but rather monetize through advertising. The result was a massive SEO and SEM industry.
The next trend is going to be AI Model Optimization and AI Model Marketing. People are already asking AI for recommendations for everything from leadership to cars to local restaurants. Brands will soon be positioning their online content, both their website and the PR, social media, and other content they can control or influence, to help them “win” in certain types of prompts. Likewise, search results will inevitably have ads shown alongside them.
Where it gets interesting is when the commerce mixes with the models. Today’s LLM models are static for everyone, just like the original search engines were static for everyone. However, today's search engines customize results based on prior search history and other factors (e.g., location). As the cost of models comes down and efficiency of models goes up, we can imagine models that dynamically adjust weights slightly based on your prior patterns. “What patterns?” you ask. It could be prior prompts, location, etc., but it undoubtedly also be based on prior ad clicks.
You can even imagine real time bidding so that the winner gets to overlay some weight vector or other content on the model. It may not be on the model itself, but perhaps a RAG-like overlay. For example, the first run of the system could produce results, but then a vector (at the model level) or content (at the prompt level) is mixed in and the initial output re-run through the model again or modified some other way. The result is a bias towards that additional weighting or content. And for the right price you can bias that result towards your product!
For the most part it will be similar to search engines today. There is, however, at least one subtle but important difference. (“At least,” because there are probably more.) When the search engine results come up the bias is clear. Sponsored results are listed as such (in many search engines), and ads are clearly marked. The nature of the output of a search engine is a list and it’s easy to scan down to see lower ranked options; even so we know there’s a huge advantage to be on page one and even the top part of page one. With output from an LLM the ads on the side may be clearly marked. The bias in the algorithm will not be so clear. Likewise, unless you asked the output to list the “top 10 X” the prose nature of the output may make it less convenient to look past the first few options listed.
Solutions like DuckDuckGo allow for less biased search results (it still is impacted by SEO choices of websites). Hopefully we’ll have similar options with AI once advertising becomes pervasive.
This also means we’ll see a massive shift in ad spend. We saw radio advertising dollars move to TV; and then TV dollars move to the web at a faster rate. We’ll see SEM dollars shift to AIMM. As before, it upends the status quo and the future winners and losers at the company and the employee level may not be the ones on top today. Plan appropriately.
Finally, let’s not forget everyone’s favorite monetization: selling consumer data. LLM customers were worried from the outset about proprietary customer data leaving their walls. Many LLMs responded by letting them run locally, inside the corporate firewall. For consumers it’s a different story. We may have some data privacy protection, but just like cookies let companies “anonymously” track users and their search results or other actions, similar techniques will be employed with LLMs. (I put the word anonymous in quotes because it’s not that hard to break the anonymity in many cases.)
The good news and bad news are that we know what’s coming in the world of LLMs. We've seen this story before. That’s good in that it means we know how to plan for it and opportunistic companies and workers can set themselves up for success. It’s bad in that it’s the consumer who inevitably loses in this story.
It’s critical to learn about corporate culture before you accept a job offer but it can be awkward to raise such questions. Learn what to ask and how to ask it to avoid landing yourself in a bad situation.
Investing just a few hours per year will help you focus and advance in your career.
Groups with a high barrier to entry and high trust are often the most valuable groups to join.
