Tap Water, AI Progress, Fancy Targeting, Marketing Attribution

After two months of traveling, I’m finally back at home in Denmark.

A few thoughts and observations that have been on my mind lately.

☠️ Tap Water

I’ve started to take the idea that “tap water is bad” more seriously. Justin Mares keeps talking about it. There is an increasing number of reports of weird stuff (pesticides etc.) found in tap water in most countries. There are also more and more studies that link the weird stuff to hormonal changes.

None of it is conclusive yet. But big issue with tap water testing is that you can only detect what you test for. So “safe” tap water can contain high concentrations of weird stuff simply because no one is doing the corresponding tests. Also the long term effects of many pesticides is simply not known yet and the current (kinda insane) approach is that stuff is considered harmless unless proven otherwise.

I bought a ZeroWater filter and an AquaBliss filter for my shower now. My girlfriend and I do feel better since we made the switch. Could be placebo, of course. But I’m testing lots of stuff, using my “perceived energy level” as metric, and most of it has no effect.

Avoiding potentially harmful chemicals is an infinite rabbit hole. Weird stuff is everywhere once you start looking. Shower gel, clothing, food, food packaging.

But drinking and showering in clean water plus eating organic food seems like a sensible 80/20 approach to me.

🤖 AI Progress

It seems like AI progress is now only happening incrementally and no longer in big jumps.

If anything it feels like many models are getting worse. Probably because the companies behind them need to save money. (My current go-to model is Sonnet-3.5.)

We do use LLMs at Sales.co quite a bit (writing copy, personalisations, labeling) and would love to automate more. But most of the time, it turned out that a few if/else rules do work better.

In every instance, we’re using AI, we still have a human in the loop. LLMs by their very nature are unpredictable. And it’s impossible to catch all edge cases programmatically, no matter how many AI-checks-itself steps you implement.

Imo everyone selling “fully autonomous AI agents” is a scammer. And this will not change anytime soon.

🚨 Fancy Targeting

One of the best posts I’ve read recently is "Anti-personalization: The best ad for one, is the best ad for all” by Samuel Brealey.

It articulates perfectly one of the big lessons I learned after running cold email campaigns for hundreds of companies:

Focus on figuring out messaging plus an offer that resonates, instead of on fancy targeting.

Fancy targeting is a waste of time and resources. Everyone who claims otherwise is either a charlatan or doesn’t understand statistics.

The problem is that fancy targeting campaigns always sound great in theory. They are awesome to sell. Companies specifically demand them.

“Let’s only target companies that recently started using Birdeye and have an open job post for a social media role so we can pitch our recruiting services.”

But when you run a prober test, it always turns out these campaigns do not perform better and often even underperform boring campaigns focused on the companies customer profile.

You simply cannot predict what’s going on inside a company from the outside.

In addition, once you zoom out you quickly realise these small-scale campaigns with fancy targeting do not matter. The number of companies reached is tiny. And you would contact the exact same companies in your boring campaigns anyway.

Sure, your messaging might be a bit more targeted in the smaller campaigns. But all this does, is produce more replies along the lines of “this is a really good cold email!” but not more deals closed.

Companies do know their situation better you ever will. And if you find winning a copy + offer, they will be interested no matter what. The winning copy + offer outperforms all alternatives regardless of targeting.

(There are a few, boring exceptions, of course, where campaigns with slightly more fancy targeting outperform boring campaigns.

📈 Attribution

Another great post I read recently is “Attribution is Dying. Clicks are Dying. Marketing is Going Back to the 20th Century” by Rand Fishkin.

Attribution is becoming increasingly more difficult and inaccurate.

Technological changes are a big factor. But at least just as important are changes in human behavior.

Most people nowadays are sophisticated when it comes to how they use the internet. Few people click on ads or reply to emails from strangers.

When something catches their attention, they do their own research first, for example, by typing the company name into Google.

Your alternatives are:

• build more and more complex models and technology to try to track what you can

• stop using fancy attribution technology and start running old-school lift-based tests

Method 1 fails when people do their own research no matter how fancy your model. There is simply no way to track that your cold email or ad prompted someone to search for the company on Google.

The best you can do is simply asking people “How did you hear about us?” when they book a call, which relies on people’s far from perfect memory.

Lift-based tests are slow but pretty accurate when done properly. For example, when we stopped our own outbound campaigns, we saw a significant drop in “inbound” leads.

Unfortunately, just as with fancy targeting, fancy attribution is simply more sexy and easier to sell.

---

Talk soon,

Jakob