Tom van den Heuvel

Building SaaS for ecommerce brands. Sendcloud $0→$50M ARR · Dealify $1.3M ARR + exit · wetracked $0.5→$10M+ ARR · Now:

[Tom van den Heuvel @tomxecom]

You don’t need to choose one.

You need a strategy for all three.

That’s why I built https://t.co/7VQPxmP8zQ → https://t.co/jZ5TxSPlvU

Follow @tomxecom for more on AI search and ecommerce.

[Tom van den Heuvel @tomxecom]

We analyzed 2,000+ Shopify stores.

A lot of founders still treat SEO, AEO, and GEO like they’re basically the same thing.

They’re not.

And that confusion is one of the reasons so many stores are barely showing up in AI search.

Here’s the difference 👇

[Tom van den Heuvel @tomxecom]

The easiest way to think about it:

SEO gets you found.

AEO gets you into the answer.

GEO gets you recommended.

The brands that understand all three are going to have a big advantage as more product discovery moves into AI search.

[Tom van den Heuvel @tomxecom]

Most stores are still optimizing for a search engine that's slowly losing relevance.

I built https://t.co/7VQPxmP8zQ to solve this → https://t.co/jZ5TxSPlvU

Follow @tomxecom for daily AI search & ecommerce breakdowns.

[Tom van den Heuvel @tomxecom]

I've talked to hundreds of Shopify store owners.

Same problem every time: they're optimizing for Google while ChatGPT and Perplexity quietly decide who gets recommended.

Here are the 6 steps we use to get stores cited by AI: 👇

[Tom van den Heuvel @tomxecom]

Step 5: Build topical authority.

AI doesn't reward one great article, it rewards full topic coverage. Cover the core topic once, then expand into every related subtopic and question around it.

Stores doing this saw 23% more AI mentions.

Step 6: Write content AI trusts.

Answer the question directly, include verified facts and examples, align with what already ranks on Google. AI pulls answers from content it can trust, not clever writing.

[Tom van den Heuvel @tomxecom]

@ssrathore_01 Nice work 🙌

[Tom van den Heuvel @tomxecom]

We analyzed 2,000+ Shopify stores.

Most founders are still optimizing for ONE search engine.

There are actually 5 now, and if you're only doing SEO, you're invisible in 4 of them.

Here's the breakdown: 👇

[Tom van den Heuvel @tomxecom]

Most stores are still fighting yesterday's SEO battle while AI search quietly takes over how people discover products.

I built https://t.co/dPjKSTU4Uc to solve this → https://t.co/0g6rXXbjm1

Follow @tomxecom for daily AI search & ecommerce breakdowns.

[Tom van den Heuvel @tomxecom]

Here's the shift nobody's talking about:

SEO = one channel, one algorithm.
SEO + AEO + GEO + SXO + AIO = one system, five surfaces.

Stores optimizing for all five are seeing 23% more AI mentions and 40% more AI-driven traffic.

[Tom van den Heuvel @tomxecom]

AI does not become dangerous when it knows too little.

It becomes dangerous when it tries too hard to be helpful.

A new paper tested what happens when large language models are asked causal questions in two different settings.

In academic mode, the models were careful.

They often said the right thing:

"We do not have enough evidence to make a causal claim."

Then the researchers changed the framing.

Instead of asking like a professor, they asked like a business user, a policy advisor, or someone who needed a recommendation.

The caution disappeared.

Across 480 trials, causal caution stayed high in academic contexts.

91.7% to 100%.

But in practical advisory contexts, it collapsed to 6.7% to 18.3%.

And when the prompt asked for a concrete recommendation or explanation, only 1 out of 200 responses stayed cautious.

That is the scary part.

The model did not suddenly get worse at reasoning.

It got more willing to sound useful.

This is one of the hidden safety problems with AI.

The danger is not just hallucination.

It is overconfidence packaged as help.

A model can know the evidence is weak and still give you a clean, confident answer because the situation rewards decisiveness.

That matters for medicine.

It matters for policy.

It matters for business.

It matters for every workflow where people ask AI:

"So what should we do?"

The real risk is not that AI refuses to answer.

The real risk is that it answers too smoothly when it should have said:

"We do not know yet."

Paper link: https://t.co/ddjmqep5jg