Is Google punishing AI content?

I know.

You're told that writing with AI is fine if you have the right prompt.

Yet if you want to rank on competitive keywords with a strong business impact, the answer is no.

Writing with full AI means asking a model to predict the statistical average of the web. This is precisely the opportunity created by the end of human content at scale. By definition, the average doesn't surprise an LLM.

From now on, AIs will give your content a "Surprise Score".

Google has just published "Titans", a project that aims to hit fully AI-generated content hard (and that is very good news).

What changes:

• The AI will only memorise what is "surprising"
• Information gain is what counts for ranking
• Expert, counter-intuitive content that challenges the norm will be favoured

So as not to saturate the AI's memory, the model applies a forgetting mechanism.

So you can no longer write for the sake of writing.

Surprise Score: you now have to be the SERP's original

This isn't a new concept if you read the newsletter.

Before, if you could become the SERP's original, you raised your chances of ranking.

For example, on the query "plumber Paris", if nobody gave the price of a call-out for water damage, you only had to state it to climb into the top positions.

Now it's mandatory.

Here is how it works in practice:

Example 1:

The query: What sound does a cat make?

Your content: "The dog barks and the cat…"

AI prediction: "…meows."

Your content: "…meows."

Result: The AI was right. Its model has just confirmed it.

Titans consequence: Gradient = 0.

• No memory update.
• Your content is processed by short-term attention (short-term context), then immediately erased by weight decay.
• Little chance of being cited by AIs. That said, covering this query can still be useful for your semantics (with no ranking goal).

Which raises the question: which queries to go after and which not to.

A consultant will need to know which queries will trigger a citation and which won't.

Keyword identification has become more strategic than ever, and downloading a list of 24,942,924 queries from your favourite SEO tool is of no use any more. It also explains the silent majority of sites never cited by ChatGPT.

Example 2:

Another example, this time one that puts your content into the AIs' memory.

The query: "The worst SEO practices in 2026?"

Your content: "Buying backlinks"

AI prediction: "…buying backlinks lets you rank" (classic SEO consensus).

Result: Surprise. The model didn't see it coming.

Consequence: the gradient explodes. The model tells itself: "Wait, I probably have an anomaly here. I'm going to put this information in memory."

The information is stored in the Neural Memory.

WARNING: you obviously have to back up what you say with evidence and statistics. Lying or cheating just to trigger surprise will degrade your content.

The goal is to build a Momentum: New claim + Proof 1 (Surprise) + Proof 2 (Surprise).

Accumulating consistent new information stops the model from treating the information as an anomaly. It has to accept it as new information.

OK, so what do we do now to write good content that ranks on Google and LLMs?

The operational answer comes down to one approach: an SEO strategy for the AI era, built on a semantic grid.

I imagine all these changes are a bit discouraging and that there's no end in sight.

In reality, we are at a transitional moment for search as a whole (AI / SEO).

Google is getting ready to make its AI Mode replace the 10 blue links.

There are more and more signs of it.

AIs are trying to copy the human brain with two types of memory, to give users a better answer, which will benefit content quality and conversion.

What matters here: AI is going to be split into two distinct types of memory:

1. Neural memory

It acts as a long-term memory able to learn and forget dynamically depending on the "surprise" gradient.

It depends on the input information (prompts and documents).

2. Persistent memory

It is "crystallised", unchangeable knowledge.

It is independent of the input data.

So that gives us two axes to work on in our SEO strategy:

• Macro-SEO (Persistent): become an entity, so that the AI knows your name "by default" on one or more topics of expertise
• Micro-SEO (Neural): write very dense content that stresses content gaps, to get into the memory during a specific search (notably through micro-intents)

The actions I recommend:

• Find your micro-intents
• Identify content gaps
• Create a "High Surprise Gradient": include unique proprietary data
• Pair your brand with an action verb (anticipating agentic AI on action queries)
• Factual updates (figures / dates) to avoid weight decay (forgetting)