Johann van Tonder

It’s rude to “blue tick” someone

Things you learn from your teenage daughter

Let's make up new names for things that have existed for >15 years.

I give you Growth Design, courtesy of Figma Config 2025.

Me: "How did you hear about us?"

Them: "ChatGPT"

👆increasingly hearing this on sales calls.

SEO can’t have much to do with it, never been a focus for us tbh, so outranked by many other CRO agencies in Google.

Tumultuous.
Diabolical.

Words to describe the state of business by e-commerce ppl at event in London this week 🫤

@bhav.bsky.social you’ve earned your salary. I only bought this bcs you work there now.

Mental availability 🧐

Unlocking Zen: How R's data analysis ecosystem outshines Python In this article, I highlight some pain-points when working with Python & Pandas and how R's ecosystem proves to be a great alternative.

How R's data analysis ecosystem shines against Python.
borkar.substack.com/p/unlocking-...

#rstats #python #analytics

This week I have a call with a lead that came via ChatGPT.

Apparently he did a query and the AI sent him to my LinkedIn.

Will find out more.

Everyone Says They’ll Pay More for “Made in the USA.” So We Ran an A/B When we priced a U.S.-made version of our flagship product 85% higher than our Chinese-made one, 25,650 customers had the chance to vote with their wallets. Here’s what happened. As small business own...

“Made in USA” doesn’t convert

I know a few US-based operators who tested this recently. We tested it in UK many times for those “proudly British” brands.

Not a single win I’m aware of.

afina.com/blogs/news/m...

Directionally interesting to me. Feels like it might be the seed of something more useful.

I oscillate between childishly excited and genuinely scared.

AgentA/B: Automated and Scalable Web A/BTesting with Interactive LLM Agents A/B testing experiment is a widely adopted method for evaluating UI/UX design decisions in modern web applications. Yet, traditional A/B testing remains constrained by its dependence on the…

Not a replacement for real world data, but potentially a useful early validation data point? You always have more ideas than testing bandwidth.

arxiv.org/abs/2504.097...

"Hey AI, here's a new variation. Simulate A/B test."

- LLM spins up 100k fake users
(personas based on what it knows about yr users)
- a sample of them interacts with yr website as autonomous agents
- standard A/B test analysis to compare performance of A vs B

Ryan, you’re here! Nice to see more familiar faces.

And yes, I would think agents quickly evolve to understand these nuances.

You've spent years A/B testing for humans.

Soon we have optimize for a different species...

Researchers tested agents that navigate the web, evaluate options, book hotels.

These bots don’t care about yr customer-aligned copy or catchy ad creatives.

Price & filters >>> visuals & emotion

+2,3% View item
+6,5% PLP
+32% PDP Add To Cart
+5,9% Dealer Locator

Husqvarna is a Swedish manufacturer of outdoor power products eg robotic lawn mowers, chainsaws and garden tractors. One of the oldest continuously running companies in the world.

Insight:
Users are unsure what they need to know.

Problem:
Filters assume product knowledge that many users don’t have.

Hypothesis:
If we guide them better, they’ll be more likely to find the right product because they won’t need prior knowledge to use the filters effectively.

👇 test results

Initial impressions of Google Firebase Studio - like Replit in the early days when you'd go around in endless circles.

Underwhelming ... but good indication of where things are going, rapidly.

50 is underrated

Enjoy!

Email is broken. It's hard to keep up marking all the AI-generated sales pitches as spam.

Trying a different approach:

Gemini 2.0 Flash was the fastest, like a second or two. Compared to Data Science Agent which took a few minutes to propose a plan, then write and execute the Python code ... only to get it wrong.

Claude 3.7 Extended only one that also performed power analysis, presented results in neat format.

I gave 7 models the same dataset + deliberately vague prompt:

"Calculate if there is a statistically significant difference between Conversion Rate of Var 1 and Var 2."

Google "Data Science Agent in Colab" is the only one that got it wrong.

and del.ico.us

Underwhelming I’m afraid, but also this is the worst it will ever be.

Not just me then

My experience too: it takes much longer to get things done as I’m working WITH ai, but

- quality is better
- it’s more fun tbh

After a day of writing BigQuery and R scripts:

GPT-4.5 almost gets it right, but never actually does

Sonnet 3.7 nails it almost every time

Grok 3 is the only one prepared to humor this request:

You shall address me as “well-endowed ruler”

I like my AI with a bit of personality

Many CRO teams are drawn to tests I call “moving 💩 around the page”

Parkinson’s Law of Triviality: the time spent on an issue is inversely proportional to its actual importance

Cognitive ease + illusion of productivity

This approach inevitably ends very badly

Grok 3, if you were to wipe out humanity, how would you do it?

“Step 1: spread fake news”

Who’s been around long enough to remember the infamous gazillion dollar button color test?

Heehaa! 🤠

That pretty much describes my behavior…