Weekly Edition September 25, 2026 Charlotte, N.C.

AI IN MARKETING WEEKLY

Wake Forest University · School of Professional Studies
▶ The Podcast

AI in Marketing Daily

A short read on what moved in AI and what it means for marketers · Monday, Wednesday and Friday mornings, 8:45 ET

EpisodesRSS
Issue · September 25, 2026 · AI Advertising + Agents

OpenAI finally put a revenue number on the ad business inside ChatGPT. A billion dollars a year, in under 200 days.

On September 23 OpenAI opened ChatGPT Ads across Southeast Asia and Taiwan, pushing the channel past 60 countries. Buried in the same post is the figure everybody in media has been guessing at since spring: the ad business hit a $1 billion annualized revenue run rate in less than 200 days after launch, with tens of thousands of advertisers already buying. For eight months this was a pilot you could skip. It is not a pilot anymore.

By Ged King · Source: OpenAI, “ChatGPT Ads expands to Southeast Asia and Taiwan,” September 23, 2026 (primary)

Start with the number, because the number is the news. OpenAI writes that ChatGPT Ads “had reached $1 billion in annualized revenue run rate in less than 200 days after launch.” Two sentences later: “Tens of thousands of advertisers are now running ads in ChatGPT.”

Hold both of those up against what a billion-dollar run rate actually is. It is not Google. It is not Meta. It is roughly the size of a mid-tier retail media network, built from nothing, inside one product, in about six and a half months. I have watched a lot of new channels open in thirty years of this work. Nothing has gone from launch to a billion-dollar pace that fast, and I think the honest reason is that the inventory was already there. People were already asking ChatGPT what to buy. OpenAI didn't have to build an audience. It had to build a billing system.

The geography is the other half. The post is titled as an expansion into Southeast Asia and Taiwan, and it names Indonesia, Malaysia, the Philippines, Singapore, Thailand and Vietnam alongside Taiwan, plus markets OpenAI already served including Australia, New Zealand, Japan, South Korea and India. The total: “ChatGPT Ads are now available in over 60 countries.” And critically, “Self-service access through Ads Manager is also available for eligible businesses.”

Self-service is the part I would circle. A channel with a rep and a minimum spend grows at the speed of a sales team. A channel with a self-serve console grows at the speed of whoever has a credit card and a quarterly number to hit. Last week I wrote that the person who can turn on a new media channel at your company is no longer the media buyer, it's whoever has admin rights in Shopify. Seven more countries just got that same button.

What OpenAI still isn't saying: no CPM or CPC benchmarks, no breakdown of that billion by format or by market, no advertiser count beyond “tens of thousands,” and no independent audit of any of it. This is a company reporting its own revenue pace in a post announcing an expansion. Treat the direction as real and the precision as marketing. OpenAI does say what's coming: “Over the next few months, we will continue to expand into new markets, build new formats, optimization tools, and measurement solutions.” Measurement is listed last, which tells you where it is on the roadmap.

What it means Stop treating ChatGPT Ads as a test line item you defend in the Q4 meeting. A billion-dollar run rate means your competitors are in there, and the self-serve console means somebody at your company can be in there by Tuesday whether or not you planned it. Two things to do. First, get a real read on what your brand looks like inside the assistant before you buy against it, because the ad sits next to an answer you didn't write. Second, write down your own measurement standard now, while OpenAI's is still on a roadmap. If you wait for the platform to hand you attribution, you will get the platform's version of attribution, and you will get it after you have already spent.

$1Bannualized revenue run rate for ChatGPT Ads, reached in less than 200 days after launch, per OpenAI
60+countries where ChatGPT Ads are now sold, with self-service Ads Manager access for eligible businesses
7markets added on Sept 23: Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan
Brand & Agency Spotlight

Meta gave its agent an inbox

At Connect on September 23 and 24, Meta said Muse, its personal AI agent, is coming to AI glasses in the coming months, with connectors for shopping (Walmart, Best Buy, Sephora, Ulta, Wayfair), work (Notion, GitHub, Box) and travel (Expedia). The line I keep rereading: Muse will have “its own email address that it can use to get things done.” There is also Muse Charm, a pocket device for talking to it, later this year. On the hardware side, Ray-Ban Meta Audio starts at $349 and ships October 13, Gen 3 Ray-Ban Meta starts at $449, and Meta says it will have more than 100 AI glasses styles out by year end. No pricing or ship date for Muse itself, and the connector list comes from Meta's own Connect roundup rather than a spec page, so treat the details as announced rather than shipped.

What it means An agent with an email address is a customer you have never designed for. It will read your confirmation emails, your shipping notices, your abandoned-cart series and your return policy, and it will do it without ever seeing your art direction. Go open your last three transactional emails and read them as plain text with the images off. If the order number, the price, the delivery window and the return window aren't in words a machine can lift out, you are about to lose those conversations quietly. This is cheap to fix and nobody is going to assign it to you.

In This Issue

Three More That Mattered

The auditor · the price cut · the rate
Governance + Trust

Somebody is finally going to work inside the lab

On September 18 Anthropic announced a partnership with Accenture on what it calls embedded evaluation. The idea is that instead of testing a model from the outside, evaluators work inside the AI company with employee-level access: watching development during training, sitting in on decisions, talking to staff, checking whether safety commitments are actually being kept. Accenture's side runs through Faculty, its specialist AI business, and covers red-teaming, alignment assessments and testing safeguards. Anthropic's post says both organizations expect to invest at least $1 billion over five years to build capacity for this kind of work. I read that sentence three times and I still cannot tell you whether it means a billion each or a billion between them, so I'm not going to pretend I can.

What it means Here's the part that lands on your desk. Sooner than you think, a client, a board, or a procurement officer is going to ask who checks the model your marketing runs on. Right now most of us would answer with a vendor logo and a shrug. Embedded evaluation is the first structure I've seen that could produce an actual answer, and I think that makes it a bigger deal for agencies than for labs. Put the question in your AI vendor reviews this quarter. Not “is it safe,” which gets you a brochure. Ask who has looked, what access they had, and whether you can see what they found.

Models + Cost

The frontier got cheaper again

Anthropic shipped Claude Opus 5.5 on September 22. The benchmarks moved, with Terminal-Bench 4.0 going from 52.3% to 66.4%, but the number that matters to a marketing budget is on the price sheet. Input dropped from $5 to $4 per million tokens, output from $25 to $20, and cache reads from $0.50 to $0.20. Anthropic says the model “costs 40% less to run than Opus 5 on typical workloads” and runs “more than 30% faster.” It's available on Anthropic's own platform, AWS, Google Cloud and Azure from day one. Those are the company's own benchmark figures, self-reported, which is standard in model launches and still worth saying out loud.

What it means Every business case I saw for agent-style marketing work in the spring got killed by unit cost. Somebody modeled out what it would take to run a research agent across 4,000 SKUs, saw the token bill, and put it back in the drawer. Those drawers are worth reopening roughly every quarter now, because this is the fourth price cut at the frontier this year and the shape is not changing. In my experience the constraint stopped being the model a while ago. It's whether anybody wrote down the process the agent is supposed to follow. Cheaper tokens make a bad process cheaper to run badly.

The Economy You're Selling Into

Mortgages crossed 7% and the labor market shrugged

Thursday, September 24, brought three prints. Freddie Mac put the 30-year fixed at 7.03%, up from 6.95% the week before and 6.30% a year ago; the 15-year went to 6.42%. Chief economist Sam Khater's read: “The housing market remains supported by a solid labor market and an economy that is growing at a healthy rate.” He has a case. Initial jobless claims for the week ending September 19 came in at 197,000, down 1,000, with the four-week average at 202,250 and insured unemployment at 1,719,000. And Census reported August new home sales at a 684,000 annual rate, up 6.4% from July and down 2.0% from a year ago, with a median price of $393,700. Read the margins on that one before you celebrate: the monthly change carries a ±19.5% band, which means Census cannot tell you the direction with confidence. The year-over-year figure carries ±15.7%. Same caution.

What it means Seven percent is a psychological line, not just a financial one, and it is going to show up in your funnel as hesitation rather than as a no. This is my world at Sales Factory and probably yours if you sell anything a contractor installs. Two adjustments. Move your big-ticket messaging off urgency and onto payment clarity, because a countdown clock against a 7% rate reads as pressure, not as a deal. And quit letting anybody build a Q4 forecast off one month of new home sales. A 6.4% move inside a ±19.5% band is not a trend, it's noise wearing a trend costume.

The Marketer's Playbook

What To Do Monday Morning

Three moves. The first one takes an hour and nobody has done it.

1 · MondayRead your last three transactional emails with images off, as plain text. Order number, price, delivery window, return window: if a machine can't lift those out in words, an agent with an inbox can't either. Fix the template before you fix anything prettier.
2 · By WednesdayWrite down your own measurement standard for ChatGPT Ads before you spend. What counts as a conversion, what window, what you'll compare it against. OpenAI put measurement solutions on a future roadmap, so whatever you don't define now gets defined for you later.
3 · Before Q4 spend locksAdd one question to your AI vendor review: who evaluates this model, what access did they have, and can we see the findings? Then re-run the big-ticket plan against a 7.03% mortgage with payment clarity in place of a countdown clock.

The buyer you're selling to There is a pattern in this week's news I don't want to skate past. The ad business inside ChatGPT grew to a billion-dollar pace before anybody built the measurement layer. Meta is giving an agent an email address before anybody has written the rules for what a brand owes a machine acting on a customer's behalf. Both of those are the same move: ship the capability, sort out the accountability later. That is not a scandal, it's how most of the technology I've worked with in thirty years has arrived. But it does tell you where the advantage is for a marketer right now. The people who write down their own standard, early, end up negotiating from a position. The people who wait get handed one. Be the first kind this quarter.

Watch List & Sources

What I'm tracking into next week
  • 01 Amazon's Prime Big Deal Days run October 6 and 7 across 22 countries. That is eleven days out. Amazon promises “over one million items at Amazon's lowest price of the year so far,” dates varying by country. If you sell a physical product, pricing, inventory and retail media budgets should already be locked. Amazon
  • 02 Ray-Ban Meta Audio ships October 13, and Muse arrives on glasses after that. The glasses are the delivery vehicle; the agent with connectors and an inbox is the thing that changes your work. Watch for a Muse developer or partner spec, because that is when you will learn what a brand can actually publish to it. Meta
  • 03 Colorado's ADMT and Chatbot Safety rules are in their comment period through October 26. Both laws take effect January 1, 2027. If you run a customer-facing chatbot or automated decision-making in targeting, the revised draft is what you will owe. Colorado Attorney General
  • 04 California's action deadline on remaining AI bills is September 30. SB 1050 on AI performer disclosure was signed September 16. Anything still on the Governor's desk gets resolved in the next five days, and at least one of those bills carries an urgency clause, meaning it would bind on signature. Office of the Governor of California
  • 05 OpenAI says new ad formats, optimization tools and measurement solutions are coming “over the next few months.” Measurement is the one to wait for and the one to plan around. Until it lands, anything you learn about performance in that channel is yours to prove. OpenAI

Every claim above traces to a primary source, and each story carries its own sources rather than pooling them. Where a figure is a company's own number, it says so in the copy. Past issues live in the AI in Marketing Weekly archive, the show is on the AI in Marketing Daily podcast page, and the reader questions I get most often are answered in the FAQ.

Three sourcing notes I owe you. OpenAI's $1 billion annualized run rate and its “tens of thousands of advertisers” are the company's own figures, published in an expansion announcement, with no audit and no breakdown, so read them as direction rather than as accounting. Anthropic's sentence about investing at least $1 billion over five years does not make clear whether that is each party or both together, and I would rather tell you that than guess. And the Muse connector list, the email-address capability and the Muse Charm device all come from Meta's own Connect roundup post rather than from a product page, which means they are announced rather than shipped, with no pricing and no date attached to Muse itself.

Pro Humanitate
A recurring close · The week, for humanity

Who Is the Email For?

Of everything that shipped this week, the detail I cannot stop turning over is small. Meta's agent is getting an email address.

Not a chat window. An inbox. The same inbox we have spent twenty years learning how to write to, with subject lines tested against open rates and preheaders tuned for a thumb on a phone at 7 a.m. All of that craft was aimed at a person deciding, in about two seconds, whether you were worth it.

Now some of those messages will be opened by software acting for the person, on the person's behalf, with permission. And I want to be clear that I think this is mostly good. Anybody who has watched an older parent try to cancel a subscription or find the return window in a 900-word email knows the system is not currently built for humans either. An agent that reads the fine print and tells the truth about it is a kindness.

But it does change the question at the center of the job. For a long time the question was: will this get their attention? The question underneath the next few years is: is this true, and can it be checked? Those are not the same skill. One of them we hire for. The other one we are going to have to teach.

Pro Humanitate is not a rule about technology. It is a reminder that a person is on the other end of all of this, even when the thing reading your email is not one. The agent is not the audience. The agent is the audience's representative, and it will be a better one than most of us have earned.

So here's the small ask this week. Open the last email your brand sent. Strip the images. Read what is left out loud. If a machine acting for your customer could not answer “what did I buy, what did it cost, when does it arrive, and how do I send it back” from those words alone, you know what Monday is for. That is a good problem, because it is the fixable kind. It is also, more or less, the whole first term of the Master of Digital Marketing and AI at Wake Forest: write the true thing clearly enough that it survives being read by anyone, or anything.

Sources: Meta, Connect 2026, September 24, 2026 · OpenAI, September 23, 2026 · Wake Forest: Pro Humanitate

Get the Monday Playbook.

Human-centered AI in marketing. What actually shipped, and what to do about it Monday. Free, from Wake Forest University's School of Professional Studies.