I Went on Holiday. The Platform Didn't.

I came home from holidays and opened a platform one of my clients runs on. Things had moved. Settings I rely on were gone. Nobody told me.

My AI didn’t know either It gave me confident advice anyway.

Meta quietly pulled placement exclusions out of ad sets. Advertisers started reporting them missing on August 25. Meta published no announcement. Value rules are the replacement, and they only move bids from plus 1,000% down to minus 90%. You can make a placement expensive. You can no longer switch it off. If you had brand safety exclusions sitting at ad set level, they just turned into a discount. Account-level Placement Controls still work, so go rebuild them there.

Google Ads is dropping campaign-level language targeting. From September, Search campaigns and the Search Network part of Performance Max match on the language of your ad copy, not your setting. Google says no action is required. That is true right up until you run one account across two languages. Go and read your ad text before the system reads it for you.

The AI Max migration finishes this month. Dynamic Search Ads, automatically created assets and campaign-level broad match are being upgraded, and Google expects every eligible campaign done by the end of September. Doing nothing is not neutral. Doing nothing is consent.

ChatGPT ads went live across 31 European markets on August 24. That is six months after the US pilot, and it takes ChatGPT ads to 35 countries. They show on Free and Go plans only, sit outside the model's answer, and run without personalisation in Europe to stay inside GDPR. Buying is through OpenAI's ads team and partners for now, with self-serve coming later.

The search shift finally has numbers on it. Organic search sessions fell 10.5% where AI Overviews are common. Referrals from AI platforms rose 163%. Your traffic is not dying. It is moving house. Give AI referrals their own line in your report this week.

I came back from a holiday and started my re-entry pass on a client platform.

Multiple things had changed while I was gone. Not cosmetic. Real changes to how the thing works.

So I did the normal thing. I asked my AI what it sees.

The answer came back fast, clean and sure of itself. It walked me through a setting. It told me where to click. It told me what to change.

The screen it described doesn’t exist anymore.

I only knew that because I’ve been doing this for 15 years. I’ve clicked that button a thousand times. I knew it was gone. If I had taken the advice, I would have spent an hour hunting for something that had been removed, then made a change that no longer does what it used to do.

This is not a bug. This is how the tool works. Your AI's knowledge has a date stamped on it. Platforms do not. The gap between those two things is where confident wrong advice lives, and it gets wider every week you are not looking.

Last month I wrote about catching AI errors inside my own data. This is the other half. That was the AI being wrong about numbers I gave it. This is the AI being wrong about the world it cannot truly see.

So I built a gate. Nothing gets advice out of the model until it tells me what it is guessing.

Then I ran the same test again with the gate on. The AI stopped describing menus. It started asking me what was on my screen. The answers got shorter and far more useful.

What worked: Forcing the CONFIRMED and ASSUMED labels made the guessing visible. Once the model had to admit which parts were memory, it mostly stopped inventing interfaces and started asking. Pasting the live screen and telling it plainly that my screen beats its training killed the arguing.

What didn't work: It is slower. You trade a fast wrong answer for two or three rounds of back and forth before you get anything. Even with a screenshot pasted in, the model still filled in the edges it could not see. And here is the honest limit: the gate only helps if you know enough to spot the wrong item when it is flagged.

It’s a seatbelt. It’s not a driver.

That is the part I keep coming back to. You cannot sit on the sidelines and expect things to grow without input. Growth needs competency, and competency means people.

The AI did not catch this. Years of clicking that button caught it.

If you are the one lying awake thinking AI is going to take it all away from you, here is what to do with that feeling. Go and be the person in the room who knows what changed.

Run a re-entry audit. It takes about 20 minutes and you should do it any time you have been out for more than a few days. Holiday, launch week, sick week, family stuff. Any gap.

Open each platform yourself. Not the AI. You. Find the What's New or announcements page and read the last month. Then open the two or three screens you touch most often and look at them properly. Write down what moved in your own words. Three lines is plenty.

Then make that list the first message in the thread, before you ask your question.

You are not competing with the machine here. You are feeding it the one thing it cannot get on its own: what is true right now, in your account, that nobody has written down yet. That is your job. That has always been your job.

Thinkers will always be the most important people in the room.

Try this today. Pick the one platform you make money on. Open it’s What's New page, read the last 30 days, and write three lines on what changed. Paste those three lines at the top of your next AI thread before you ask it anything.

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.

www.jamesklobasa.com