The Briefing
1. Google is migrating your Search campaigns to AI Max whether you like it or not. Through September, any campaign running campaign-level Broad Match or automatically created assets gets upgraded in place. There is no opt-out button. Your only choice was to switch those features off before September 1, or migrate on your own terms first.
2. Meta is taking placement control off the table. You can no longer hand-pick certain placements at the ad set level. Meta decides where your ad shows based on the objective you chose. At the same time, the Meta AI assistant inside Ads Manager is now open to every advertiser and agency worldwide.
3. ChatGPT Ads went self-serve and it is moving fast. OpenAI reported its ad business hit a $1 billion annualised run rate in under 200 days. Self-service buying through Ads Manager is now live across 31 European markets plus India and MENA, with product feeds and custom audiences. A third major ad auction just opened.
4. Google is folding Data Manager into Analytics and DV360. Enhanced conversions are rolling into both. Google claims an 11% average lift in Search conversions against standard conversion imports, and a 26% incremental ROAS gain. Treat the numbers as the vendor's. Treat the direction as real: your measurement plumbing is being centralised too.
Notice the pattern. Targeting, placement, creative, measurement. Four dials. All four are being handed to a model this month.
The Build
The migration notice landed on a Tuesday. I spent twenty minutes hunting for the opt-out. There isn't one.
Then the Meta placement change. Then I read that ChatGPT Ads had crossed a billion.
I've spent 15 years earning my keep on those dials. Picking the audience. Picking the placement. Writing the ad. This month a lot of that moved behind glass.
So I asked a blunt question. What is still mine? (After all, I secretly run my own products in a very obscure niche, and I like it paying me)
The answer is the offer. And more precisely, what happens in the sixty seconds after someone says yes.
No algorithm sets your order bump. No model decides your one-time offer. Average order value is the last big dial with my hand still on it, and I had been ignoring it while I fussed over match types.
So I pulled a file out of my own library: upsell-downsell-scripting-1.md. It's the OTO framework built off Russell Brunson's work. Order bump, one-time offer, downsell, thank-you page. Four slots, one rule each.
I turned it into a working prompt and ran my own offer through it.
You are building a post-purchase offer sequence using the OTO
framework. Here is my core offer:
CORE PRODUCT: [name, price, who buys it]
THE OUTCOME THE BUYER WANTS: [what they are actually trying to achieve]
WHAT USUALLY STOPS THEM AFTER THEY BUY: [the real obstacle]
WHAT ELSE I ALREADY OWN OR COULD DELIVER: [list assets, services, time]
Build four slots. One at a time. Stop after each and wait for me.
SLOT 1 - ORDER BUMP. Priced at 10 to 20% of the core product.
Must be a one-second yes. Must make the core product work better.
SLOT 2 - THE OTO. One offer only, no menu. It must help the buyer
reach the SAME outcome as the core product, faster or with less
friction. An adjacent product is not an OTO. If your idea solves a
different problem, discard it and try again.
SLOT 3 - THE DOWNSELL. Same offer on a payment plan, or a lite
version. Not a new pitch.
SLOT 4 - THANK-YOU PAGE. One low-friction offer, or nothing.
HARD RULES, do not break these:
- Invent zero numbers. No made-up "total value" figures, no fake
testimonials, no stats I did not give you.
- Scarcity must be true. If I can offer the same deal next month,
do not call it a one-time offer.
- The decline link must be honest and pressure-free. No guilt,
no shaming, no "I'd rather stay broke" copy.
- Acknowledge the purchase they just made before pitching anything.
For each slot, give me: the offer, the price, the reason it passes
the congruence test, and the script in my voice.What worked: the congruence test did the heavy lifting. The framework has one rule that matters more than the rest: the upsell must get the buyer to the outcome of the thing they just bought, faster.
The first batch of ideas Claude gave me were adjacent products. Useful things. Things I could sell. But they solved a different problem, so they failed the test and I binned them. Naming that rule inside the prompt is what made the second batch usable.
The order bump it landed on was something I already owned and had never once offered at checkout.
What didn't work: two things, and both were the model doing what models do. First, the value stack. Unprompted, it produced a tidy "total value" figure built entirely out of air, with prices attached to bonuses I had never priced. I stripped every one.
Second, the decline link. It wrote guilt copy. The kind that makes a buyer feel small for saying no. The framework explicitly warns against high-pressure tactics, and the model walked straight past that line until I promoted it to a hard rule and quoted it back. That's the version you see in the prompt above. It took three passes to get there.
Here's what I actually took away from the week. The platforms are not stealing your job. They are taking over the part of your job that was always going to be automated, and they are doing it faster than most people planned for.
What they cannot touch is the shape of your offer, the order you present it in, and whether the second thing you sell genuinely helps the buyer with the first.
That work is still manual. That work is still yours.
The Play
If you're a founder or marketer who just watched a platform take a control panel off you this month, stop trying to win it back. Go where you still have a vote.
Open your checkout. Look at your best-selling product. If there is no order bump on that page, you are leaving money on the table every single day, and no amount of targeting work will make that up.
Start with one question, and make your AI answer it properly:
My best seller is [product] at [price]. The buyer's real goal is
[outcome]. After they buy, the thing that usually slows them down
is [obstacle].
Give me 5 order bump ideas priced between 10% and 20% of the core
price. Each one must make the core product work better or faster.
Then score each idea 1-10 on ONE thing only: does it get this
buyer to [outcome] sooner? Anything under 8, throw out and explain
why it failed.
Do not invent prices, values or results. Use only what I gave you.That last line matters more than the rest. Every model I've tested will hand you an invented "$4,000 value" if you let it. Take the scoring seriously and you'll usually find your bump is something you already own and forgot about.
Try this today. Run that prompt on your single best-selling product, pick the one idea that scores highest, and add it to your checkout page as a simple checkbox before you touch another ad setting.
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.