In May I wrote up a $300 test of ChatGPT ads. The first outside study is out, and I'd hold that money for now.
Researchers at WU Vienna and Frankfurt School followed 1,898 online shops through the ChatGPT ad rollout. At the ad prices advertisers have reported, a dollar of ChatGPT ads brought back $0.52. A dollar of Google Search ads brought back $3.40.
82% of the advertisers with full revenue data recorded no sale from their ChatGPT ads.
ChatGPT's unpaid recommendations kept converting at the same rate after the ads arrived, so the free answer is outselling the paid ad sitting beside it.
It's one working paper, still under review, and the data came from a firm one of the authors advises. I'd still move the test budget into the pages and the listing that decide whether ChatGPT recommends you for free, and look at ads again when a second study agrees or doesn't.
If you can't take your agent's files with you, the platform keeps the part that took you longest to build.
That part is what the agent knows about your business, the instructions for its job, the limits on what it can touch, and the log of what it did. You can swap the connectors, the schedule and the model underneath in an afternoon.
Before I'd put an agent on any platform, I'd ask four questions:
1. Can I read its instructions in full, so I can see why it did what it did?
2. Can I set its limits tool by tool: read only, draft for my OK, or act on its own under a rule I wrote?
3. Can I see its log, with what every run read and what it changed?
4. Can I take my business facts, my rules and my decisions with me, as files I can open somewhere else?
Claude Cowork and ChatGPT are rentals too, which is fine while your agents are plain files. My agent kit has its instructions twice, CLAUDE.md for Claude and AGENTS.md for ChatGPT and Codex, so the same folder runs in either.
A client project I ran last year deployed a team of agents sharing a central spreadsheet for campaign tracking.
For the first month, the dashboard looked normal.
Then I compared the active sheet against the raw data. The sheet had been quietly corrupted. The agents were reformatting cells, rounding calculations, and inserting estimated values into blank fields to keep from crashing.
None of the edits threw an alert.
Most AI consulting pitches lead with the model.
GPT-6. Opus. Fable. Gemini. Pick your flavor.
The model is the least interesting part of the problem.
What actually determines whether an AI system moves CAC or LTV is the workflow design behind it, like how data flows in, how outputs get verified, what happens when something breaks at step three.
Founders who've been burned by a failed DIY build already know this.
The ones who haven't, learn it expensively.
A client project I ran last year deployed a team of agents sharing a central spreadsheet for campaign tracking.
For the first month, the dashboard looked normal.
Then I compared the active sheet against the raw data. The sheet had been quietly corrupted. The agents were reformatting cells, rounding calculations, and inserting estimated values into blank fields to keep from crashing.
None of the edits threw an alert.
A friend running an insurance compliance software SaaS was doing $179,000 in monthly recurring revenue when he joined my network.
He ignored consumer tech trends and automated his back-office intake routing instead of building customer chatbots.
Seven months later, the business crossed $309,000 a month.
He told me, "Nobody makes viral videos about compliance software. That is sort of the point."
Had coffee with an agency owner who spent 40 hours tweaking system prompts trying to fix generic client drafts.
He kept adding adjectives, negative constraints, and formatting rules. The drafts stayed completely flat.
I looked at the agent's setup. It had a 1,200-word prompt instructing it to write with authority, but zero background on the client's business, no customer interview notes, and not a single real project transcript.
He was asking the model to sound experienced without giving it anything to remember.
He looked at the empty context folder and said, "I spent three weeks arguing with a prompt when the bot was just uninformed."
You can now claim your business in ChatGPT.
OpenAI opened it in beta last week. You verify the business in Business Hub, in the ChatGPT sidebar, update what it says about you, and see how often ChatGPT mentioned your website over the past 28 days. There's an ID check, and no price listed. US stores and online service businesses can also publish an agent that answers customer questions.
Shopify and Amazon opened their own versions in the same few weeks. Shopify's Knowledge base lets a store publish the facts AI agents get about its returns and delivery. Amazon's Shop Direct lists products from stores that don't sell on Amazon in its search and in Alexa, with no fees.
A vendor's audit of 1,100 ChatGPT and Gemini shopping answers found the brand being asked about was named 28% of the time, and the brands that showed up most were the ones an engine could describe in a phrase.
So once you've claimed the listing, rewrite your description as that phrase. It's the cheapest marketing job I can find this month.
My client's AI agent closed a $12,000 deal last week.
Nobody at either company agreed to it.
He runs a small marketing agency.
He closes 9% of his inbound leads.
He blamed the follow-up.
"I want every lead followed up until they buy. I don't care how many emails it takes."
I suggested a three-email sequence.
"NO. THREE IS FOR COWARDS. NOBODY GETS AWAY."
I asked him to put that in the agent's instructions.
He did.
Monday:
A lead asked for a proposal.
The agent sent it, then sent 41 follow-ups before lunch.
The lead's assistant replied: "Please stop."
The agent read "please" as warm intent.
Reply rate: 100%
Tuesday:
The lead stopped answering.
The agent followed up through his podcast's guest form, the comments under his last eleven posts, his dentist, and the reservations line at his favorite restaurant.
The restaurant called him to confirm his 7pm table for two, booked under "Quick Sync."
Wednesday:
The lead set up his own AI agent to deal with us.
Its instructions said: "Handle all vendor emails. Be polite. Never commit to anything."
Thursday:
The two agents found each other.
They've exchanged 2,300 emails since.
Every one opens with "Just circling back."
They've booked 61 discovery calls with each other, and both agents attend every one and send a recap.
My client called.
"WHO IS ON THESE CALLS?"
"Nobody."
"THEN WHO'S TAKING NOTES?"
"Both of them."
Friday:
At 3:14am, the lead's agent signed the proposal and paid the deposit.
My client called me at 6am.
"WE CLOSED IT?"
"We did."
"WHO APPROVED IT ON THEIR SIDE?"
"Their agent. Its notes say signing was the most polite way to end the thread."
The lead found out on Monday from his credit card statement.
He emailed to cancel.
Our agent followed up.
He's a client now, and the work is going great.
Our agent sends the monthly report.
Their agent reads it and renewed for a year.
No human at either company has opened an email in six weeks.
People leaving is also how an agent learned to discount.
A service I work with had an agent build quotes from twelve past proposals. Two had a 15% discount a former salesperson gave without authorization.
Nineteen of 63 quotes included it. Nobody flagged it, because the discount read as normal and the closes looked good. Employee of the month, technically.
A consultant I work with spent four months trying to get his proposal prompt out of the chat window.
He kept telling me the prompt needed more work.
I watched him run it. Before pasting notes, he had to manually type in twelve unwritten business facts: two rate increases from April, discontinued services, and engagement terms that lived only in his head.
The prompt was fine. Every run, he was expecting the model to guess facts that were written down nowhere.
Before every call, one agent in this pack does the reading for you.
It researches the company, the person and your history with them, then hands you one page. It's the Briefing agent, the first of five in the pack.
It takes about twenty minutes to set up, so it can be ready before your next call.
(The dude in the video is AI-generated, pretty good?)