AI in Marketing Weekly

Issue · September 11, 2026 · Retail Media + Media Buying

Amazon started selling ads inside ChatGPT on Thursday. The chatbot's ad business is 200 days old.

On September 10, Amazon Ads opened ChatGPT inventory to a select group of U.S. advertisers through the Amazon DSP, with Delta Vacations first in the door. It's a managed-service pilot, priced on clicks or impressions, and OpenAI keeps control of where each ad lands. The company selling the placements booked close to $70 billion in ads last year. The company serving them started running ads in February and already claims a $1 billion annualized run rate.

By Ged King · Sources: Digiday, “Amazon Ads brings advertisers to ChatGPT, extending its supply-chasing streak” (primary) · Marketing Dive, “Amazon pilots ad services in ChatGPT: What marketers need to know” (primary)

Here's the plain version. If you already buy through the Amazon DSP, Amazon's team can now set up and run your campaign inside the ChatGPT app. The ads show up as text and image units under the chatbot's answer, labeled sponsored, on the free tier and the cheaper Go plan. You get aggregated numbers back: impressions, clicks, cost per result, CPM, CPC. That's the whole offer for now, and it's U.S.-only.

Why I think this is the week's story: it puts the two most useful data sets in advertising on the same buy button. Amazon knows what you bought. ChatGPT knows what you asked about at eleven o'clock last night. Delta Vacations is a smart first test because nobody buys a trip to Cancun on Amazon, but plenty of people plan one in a chat window. If a conversational placement can move a product Amazon doesn't sell, Amazon's pitch to every travel, finance, and auto brand writes itself.

It's also the same play Amazon has run for eighteen months. Netflix, Roku, Spotify, Disney, now ChatGPT: lock down the inventory buyers want most, charge close to nothing on programmatic guaranteed deals where The Trade Desk has historically taken 15 to 20 percent, and let the shopping data do the selling. OpenAI reportedly talked to The Trade Desk first and chose Amazon anyway.

Now the part I'd underline. The measurement is thin, and OpenAI decides placement, not you. eMarketer's Nate Elliott said this week that OpenAI is still building the basics of an ad sales operation: team, tech, formats, pricing. And Amazon is expanding into ChatGPT ten days after the FTC accused it of hidden auction surcharges. None of that means don't test. It means test the way an engineer would, with a question you can answer and a holdout you can point to.

What it means If you buy through the Amazon DSP, ask your rep this week whether you qualify for the pilot. Before you say yes, write down the one question the test would answer, like “does a chat placement move a product we don't sell on Amazon?” If you can't write that sentence, you're not ready to spend the money.

$1BChatGPT Ads annualized run rate, reached in under 200 days

~$70BAmazon's ad revenue last year, the third-largest platform

12% → 27%share of U.S. ad spend run through AI campaign types, 2026 to 2030 (Madison & Wall)

Brand & Agency Spotlight

Klaviyo took the walls off its own product

On September 9 at K:BOS in Boston, Klaviyo announced it's going headless: more than 260 MCP tools and 490 APIs now callable from Claude, ChatGPT, or whatever agent a team already uses, so a marketer or an agent can read data, write a campaign, and send it without opening Klaviyo at all. It also added SQL inside its data platform, so a plain-English question returns an answer plus the query that produced it, and said its Customer Agent has exchanged 1.3 million messages with shoppers, with voice coming next. Most software vendors treat the login screen as the moat. Klaviyo just bet the moat is the data underneath.

What it means Ask your CRM or email vendor one question this month: can my team run it from the AI tools we already live in? If the answer is “on the roadmap,” you now have a named competitor who shipped it.

In This Issue

Three More That Mattered

The forecast · the bots · the audit log

Marketing Data + Media

One in eight U.S. ad dollars now runs through an AI campaign type

The IAB raised its 2026 U.S. ad-growth forecast from 9.5% to 12.3% this week, crediting the Olympics and the World Cup. The number underneath it matters more: Madison & Wall estimates 12% of U.S. ad spend now flows through automated campaign types like Advantage+ and Performance Max, up from 2% in 2023, and expects 27% by 2030. Roughly 30% of Google search spend already uses AI Max or PMax; Meta's Advantage+ is on pace for $75 billion this year.

What it means Pull your own version of that number: what share of last quarter's media ran through an AI campaign type? If nobody on your team can answer by Friday, that's the finding.

E-Commerce + CX

More than half the web's visitors aren't people anymore

Cloudflare says bots and automated agents now make up more than half of all web requests. One agency told Digiday its e-commerce clients saw bot traffic rise 80% year over year, and that tightening retargeting to exclude suspect visits pushed CPMs up 20%. The catch: AI agents add to carts and sign up for newsletters, so the old bounce filters miss them. (Digiday) One pillow brand said AI-assistant referrals convert several times better than its site average.

What it means Ask your analytics lead to split “AI referral” from “bot” in next month's report. Classify first, block second. Some of that traffic is your best customer with a robot doing the typing.

Design Thinking + Org Strategy

Agencies are building audit logs so their AI agents can't run up the bill

Gartner found 56% of companies deploy AI with no clear usage policy and expects 60% of AI users to face cost overruns from untracked usage. So agencies are building the tracking themselves: Rise logs every change its media-buying agent makes through PubMatic's audit tool, PMG caps daily token use per person, and Dept routes every prompt through a gateway after one staffer burned 1.5 million tokens in a day. (Digiday) Dept's rule: every deliverable has an accountable human.

What it means Ask your agency for the decision log on any campaign an agent touched. If they can't produce one, no agent-run campaigns on your account until they can.

The Marketer's Playbook

What To Do Monday Morning

Three moves with dates on them. None of them costs media money yet.

1 · This weekAsk your Amazon Ads rep whether your account qualifies for the ChatGPT pilot and exactly what reporting you'd get back. Write the one question a test would answer before you agree to spend a dollar.

2 · By next FridayPull the share of Q3 media that ran through AI campaign types: Advantage+, Performance Max, AI Max. Put the number in front of your CFO with one sentence on what you can and can't see inside it.

3 · Before Q4 budgets lockRequire a decision log and a holdout for every agent-run campaign, yours or your agency's. No log, no agent. Write it into the next scope of work.

The buyer you're selling to Freddie Mac put the 30-year mortgage at 6.76% this week, up from 6.71%. The latest CPI read holds at 3.4%, with the August print due at 8:30 this morning. Michigan's final August sentiment landed at 51.7, down 6% on the month and 11% below a year ago, and the September preliminary comes at 10:00 today. The customer your ChatGPT ad is about to meet is still paying more to borrow and still bracing for prices, and no amount of placement precision changes what's in the checking account.

Watch List & Sources

What I'm tracking into next week

01 This morning's two releases: August CPI at 8:30 and the preliminary September Michigan sentiment at 10:00. Both landed after this issue was written, and both feed the Fed's September 16 decision. BLS · Surveys of Consumers · FOMC calendar
02 Brandweek in Atlanta, September 15–17, which adds a full AI Accelerator Day this year. Watch for which brands say what they've actually shipped versus what they're piloting. Adweek
03 Whether the Amazon–ChatGPT pilot opens beyond a select U.S. group before Q4 spend commits, and whether either company shows third-party measurement. Digiday
04 Google's AI Max migration wraps September 30. Last week's story isn't over; I'm still waiting for any performance data on the accounts that were moved. Search Engine Land

Every claim above traces to a first-hand source. Links go to the original reporting; where I leaned on a company announcement, I say so in the text. Past issues live in the archive, and the daily show is on the podcast page.

One sourcing note I owe you: the $1 billion run rate is OpenAI's own figure, and a run rate is one month multiplied by twelve, not money booked. Klaviyo's tool counts and its 1.3 million messages are company numbers too. I couldn't independently verify any of them, which is why each one is labeled that way above.

Pro Humanitate

A recurring close · The week, for humanity

Two Data Sets, One Person

Strip the deal terms away and here's what happened on Thursday. A company that knows what you bought joined up with a company that knows what you're worried about. Your late-night question about whether the trip is affordable this year, next to your order history, in the same ad-buying system. Marketers got more powerful this week. I'd rather we said so out loud than pretended it was just another integration.

I teach my students that the customer is a person with a checking account, not a segment with a CPM. Sentiment is 11% below a year ago, the mortgage rate went up again, and the person on the other end of that ChatGPT conversation is asking for help because they're trying to decide carefully. Meeting them there with an ad isn't wrong. Meeting them there with something useless, or something that pretends to be an answer, is.

So here's my ask, and it's the same one I'd make of a media agent as of a junior planner: know why you're in the conversation. If you can't explain what the ad adds to the person's decision, don't run it. Wake Forest's motto is Pro Humanitate, for humanity. This week it means remembering there's exactly one person behind both data sets, and they're the one you answer to.

Sources: Marketing Dive · Surveys of Consumers · Wake Forest: Pro Humanitate