The Briefing
Google is migrating your Search campaigns to AI Max whether you like it or not. From September 1 to 30, campaigns still running campaign-level Broad Match or standalone Automatically Created Assets are being upgraded. Search term matching and text customisation come on by default. There is no opt-out and no route back to the old settings, so pull your search terms report this week and see what you are now paying for.
LinkedIn is throttling AI slop, not deleting it. The platform reported a 46% jump in detected inauthentic activity in the first half of 2026. Flagged posts stay up. They are suppressed so they never travel past people who already follow you. Members can now tap a "Seems like AI slop" button on any post, and the feed ranking was rebuilt to demote generic, perspective-free writing.
ChatGPT Ads hit a $1 billion annualised run rate. OpenAI is expanding into 31 more countries, and advertisers in India, Europe, the Middle East and North Africa can now buy placements straight through Ads Manager. If your buyers ask ChatGPT questions your product answers, there is now a way to be in that answer.
The EU AI Act transparency rules are live. They came into force on August 2. If your chatbot, your ad creative or your content reaches anyone inside the EU, disclosure obligations apply even if your company sits outside it. Check your site and your bot this week. It is a ten minute job now and a lawyer's job later.
The Build
Everyone reacting to the LinkedIn crackdown is trying to make their AI text sound more human…AGAIN!
BUT… Sounding human is not what LinkedIn is scoring.
LinkedIn's own words for slop are content that "can appear polished while lacking original perspective." Read that again. Polished is not the accusation. Polished is the tell.
The feed is not hunting for robots. It is hunting for writing that contains nothing only you could know.
So this week I built a filter for it, out of a skill file called composing-linkedin-stories.md. It runs seven questions the model has to ask before it’s allowed to write a word.
The last rule is the whole machine: cut any sentence another person in my industry could have written word for word.
Then I wrote this newsletter. Shipped the draft. Did not run it through the filter.
Someone read it back to me and asked whether I had. I had not. So I did, and my own draft failed on three of the four rules. Bugger!
Here is what the filter pulled out of my writing:
"Early auctions are usually the cheapest they will ever be." Every ad newsletter on earth has run that line. "Do it before someone else does it for you." Padding. And the worst one, sitting in the section where I told it to be specific: "That single answer is the whole difference between content that travels and content that sits."
That last sentence is a poster. It is the exact thing my own rule bans, written by the person who wrote the rule, eleven lines under the rule.
Question seven is the one that broke me.
It asks for the one specific thing that happened to you that proves your point. My draft had no answer. I argued the point instead of proving it, and every vague sentence in the piece traced back to that empty slot. Arrrgh!
You are helping me write one LinkedIn post. Do not write yet.
First, ask me these seven questions one at a time and wait
for my answer before moving to the next:
1. What is the product, service or idea behind this post?
2. Who is the reader? Job title, industry, what they are
worried about this month.
3. What is this post for: leads, authority, hiring, traffic?
4. Paste three or four of my recent posts so you can hear
my voice.
5. What tone or phrase should you never use?
6. What do I believe about this topic that most people in
my industry do not?
7. What is the one specific thing that happened to me that
proves point 6? A moment. A date. Something that
actually happened.
If my answer to 7 is a theme and not an incident, tell me
so and ask again. Do not proceed without it.
Then build the post in this order:
- OPENER: one line that challenges an assumption my reader
holds. No question marks. No "Here's the thing."
- ARC: setup, what went wrong, what I learned. Use the
incident from question 7. Do not generalise it.
- CLOSE: the reader's next move, in one sentence.
- TAGS: 5 to 7. Two broad industry, three niche community.
Rules: my words from questions 4 to 7 carry the post. Cut
any sentence that another person in my industry could have
written word for word.
Last step, before you show me anything: run the draft back
through that final rule and quote me every sentence that
fails it.The version you are reading now is the rebuilt one. The incident in it is the one you just read: I wrote the filter and did not point it at myself.
What worked: Making the model quote the failing sentences back, rather than fix them, is the change that mattered. Seeing "Early auctions are usually the cheapest they will ever be" pulled out and named is different from having it rewritten.
You learn your own tells.
Adding the guard on question 7, where it rejects a theme and asks again, is what stopped the whole thing collapsing back into argument.
What didn't work: The tagging rule produced tags nobody searches, so I cut "one tag I could own" from the prompt entirely and write tags by hand.
Two sample posts in question four was not enough to hold a voice, so it’s three or four now, and the model still drifts back toward polish by the second draft, so I only use draft one.
And the honest failure: the prompt had no way to make me run it on my own work. A filter you forget to switch on is not a filter. That is why the last instruction now exists.
The Play
If you are a founder or marketer who posts on LinkedIn and has watched reach fall off, the platform did not stop liking you… it started scoring perspective, and posts that could belong to anyone now die inside your own network.
Open your last five LinkedIn posts. Read each one and ask one question: could a competent competitor have published this word for word with only the name swapped?
Where the answer is yes, that post was built to be suppressed.
Then take the worst one and rebuild it with the prompt above.
When you hit question seven, do not answer with a lesson. Answer with a example. A number off your own P&L, a thing a client said on a call, the moment something broke.
And when the model hands you the list of sentences that failed the final rule, read that list twice. Those are your defaults. They are what you reach for when you have nothing specific to say.
Try this today. Run your most recent LinkedIn post through the prompt's last instruction only, and count how many of your own sentences come back.
Growth Partner for founders stuck between $1M and $20M | I find the money you're sitting on and build the machine that collects it | $1B+ in client results.