The Briefing
1. Google is retiring Dynamic Search Ads. Google's updated help docs now say DSA campaigns get automatically upgraded to AI Max beginning February 2027. The new reporting adds search term and landing page views, and Google is telling advertisers to go easy on negative keywords so AI Max can find intent. If you still run DSA, you have one planning cycle to migrate on purpose instead of getting migrated by default.
2. Amazon folded its ad console and its DSP into one agent. On September 29 at unBoxed, Amazon launched Amazon Ads Agent. The old Ads Console and DSP are replaced by a chat window, and campaign types collapse to three. The new Full-Funnel Campaign spans streaming TV and sponsored placements with a $100 per day minimum. You now brief Amazon's ad system in sentences.
3. OpenAI shipped agents that work when you are not there. Dots, announced September 29, are always-on agents that run between your conversations and connect into thousands of apps. The useful read for a small business owner: AI work is moving from "answer my question" to "watch this and act." Pick one process you would genuinely trust unattended. Start there.
4. The buy button is being automated. Shopify opened its checkout to browser-based AI agents in late September. Meta's Muse agent can now complete a purchase through Shop Pay, and Amazon promptly blocked Muse from doing it on Amazon. Soon an agent will do your buyer's clicking for them. The moment of purchase gets faster and more valuable, which is exactly the moment most businesses have nothing prepared.
The Build
This week I built an upsell and downsell skill into my AI stack.
The reason is boring. I spend almost all my time on the front of a funnel. Traffic, hooks, angles, landing pages. The part that happens after someone pays usually gets whatever attention is left over. Which is minimal on the good days...
So I loaded the one-time offer framework into Claude, gave it a core product, and asked for the upsell.
It gave me a cash grab.
The writing was clean. The value stack was tidy. The price made sense. And the offer had nothing to do with what the person had just bought.
AI can’t think like a seasoned marketer…yet!
It was a second product, sold to a tired buyer, four seconds after their card cleared.
That is the default failure. AI gives you what you asked for. I asked for more revenue, so I got revenue-shaped words.
The fix was one line of constraint, written before anything else in the prompt.
HARD RULE, APPLY BEFORE ANYTHING ELSE:
The upsell must help the buyer get the result of the
thing they just bought - faster, easier, or with less
risk. If it does not, it is a different product for a
different day. Say so and stop.
CORE PRODUCT: [what they just bought, and the price]
THE RESULT THEY WANT: [the outcome they paid for]
THE NEXT OBSTACLE: [the first thing that blocks them
after they pay]
Now do this, in order:
1. Name the next obstacle in the buyer's own words.
2. Propose ONE offer that removes it. One. Not a menu.
3. Write the decline link so it states the real
trade-off honestly. No guilt, no shame, no
"I'm okay with failing."
4. Tell me which part of this offer I cannot prove yet.The second draft was a different animal. Same framework, same model, same product. The constraint changed what counted as a good answer.
Then I hit something I didn’t expect.
The classic script tells you to write the decline link as a confession: "No thanks, I'm okay with getting results slower."
The same framework, a few lines down, lists making the buyer feel bad for saying no as a mistake to avoid.
Both of those are in the file. They cannot both be right.
I picked the honest one.
My decline link now reads: "No thanks, I'll handle this part myself." It names the trade and leaves the buyer's dignity where it was.
Here’s the handle I kept for it.
A sale is not a door, it is a rung. The buyer climbed one step and is now standing on it, looking up, slightly out of breath. Your job at that second is to hand them the next rung, not sell them a different ladder.
What worked: Putting the congruence rule at the very top of the prompt, before the product details. Forcing one offer instead of a menu. And step 4, asking the model to flag what I cannot prove, which caught three invented benefit claims I would have had to strip later anyway.
What didn't work: Asking for the bump, the upsell, the downsell and the thank-you page offer in one go. The output looked complete and was mush.
Each one needs its own pass.
The downsell was the worst of them, because a downsell written in the same breath as the upsell just repeats the upsell in a smaller voice. I also had to kill every countdown timer it suggested. A timer that resets is a lie, and buyers check.
The Play
If you are an operator pouring money into traffic while the back of your checkout sits empty, this is the cheapest revenue you own. Not a new funnel. One checkbox.
Build one order bump. Price it at 10 to 20 percent of your core product. Its only job is to remove the very next obstacle a buyer hits after they pay.
My core product is [X], priced at [$Y].
The buyer wants [outcome].
Right after they pay, the first thing they have to do
is [the boring, annoying first step].
Write me an order bump that does that first step for
them, or makes it take 10 minutes instead of 3 hours.
Rules:
- Price it between 10% and 20% of [$Y].
- One benefit. One sentence. Under 25 words.
- It must read like a favour, not a second sale.
- Then list what I have to actually build or record to
deliver it, and how long that takes.That last rule is the one people skip. An order bump you cannot deliver by Friday is a note in a document, not revenue.
Try this today. Open your checkout page. Write down the first thing a buyer has to do after the payment goes through. Then price a version of "we already did that part for you." One checkbox, one sentence, live before the weekend.
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.