Google’s Semi-Secret Agent Advantage
The AI Internet is already all around us in the form of the current Everything Machines. Their owners are, very obviously, unsatisfied with their current forms and are continuing to push their development rapidly and very expensively forward. This evolution is from Answering Machines to Acting Machines, more commonly referred to as agents.
Next Best Word
Building Answering Machines, even these imperfect ones, has been an incredible task, but a pretty easy one to summarize. At their core, all these massive GenAI systems do is predict the next word that a human would type in the context of whatever summary, essay, report, or story is being asked of them. We are predictable enough as a species that, given all the text on the Internet and a multibillion dollar cloud infrastructure, most new poetry and prose can be created for us if we request it. The same goes for images and videos. The next pixel or frame can be extrapolated with increasingly high fidelity just as the next word can. We finally have our infinite monkeys on infinite typewriters.
This also means that they are implicitly judging our text against the statistical norms used across the internet. Do they like what we wrote and lift it? Or, do they not and rewrite it? Or, would they prefer to be fed structured information in the form most economic for them and not worry about our prose? (It’s that last thing in case you are wondering.)
Next Best Action
The same method and math applies to agents. Agents will have to predict the next human action to reach a similar level of consumer satisfaction as the Everything Machines enjoy today. They will happily consume your site map and some Q&A you target at them, but that’s not how their statistical models are built. Their models will be built in the same way as GenAI above: gather enough human clicks on all the world’s websites – to predict the next click. This is a task which is effectively impossible with the datasets available. Our clicks – and the order of those clicks – have not been recorded as our words, images, and videos have been. Clicktracking is currently the stuff of spies and stalkers; but we’re going to have to think of those people as early adopters. In the legitimate world, Google Navboost is by far the most comprehensive in this regard and creates a significant advantage for Google Gemini’s agents.
Google’s Advantage
Between their unique Google Analytics and Chrome data, Google Navboost uses millions of SEO “partners” to build a record or what we all click on, creating a statistical map of how humans use online properties. Those statistics are exactly the model Google needs to roll out its initial agents, which will by then be trained by us to Act just as the current Gemini has been trained by us to Answer. To drive conversions out of Gemini and its upcoming agents, we all need to participate. However, to passively hand Google the monopoly on global clicktracking is going to make them money to the exclusion of the rest of us.
Avoiding a rerun?
We all know what a Google monopoly looks like. It is in everyone else’s best interest to ensure that doesn’t happen again. The way to avoid it (and make money in the process) is to properly train Google’s Everything Machine competitors with the data ammunition they need to compete. Otherwise, we’ll find ourselves fighting (1) Yelp’s fight all over again, or (2) over the scraps Gemini leaves for us. One significant technique for arming OpenAI and later, Apple Intelligence, is to emulate the aspects of Navboost that we can execute on practically and legally. Capture our own clickstreams, de-identify them if appropriate, and feed them to the Everything Machines today. Even Apple; even with their schedule slippage. Make sure that the LLMs that deploy and control the agents can build a statistically significant map of how your users navigate your sites. Teaching three armies of action beats watching Google dominate the agenetic internet the way it did Search & Discover.