ChatGPT Ads Sponsored Agents: OpenAI Tests Ads That Talk Back, Plus 3 AI Tools

OpenAI put AI on both sides of the ChatGPT ad

A shopper spots a dining table in a ChatGPT ad. She has questions. Will it fit the room, how many people does it seat, how do you care for the finish? That is OpenAI's own example, and until September 16 the way to get those answers was to click out to the store and go looking.

Now there is a second path, for the moment only inside a test. OpenAI's post of September 16, 2026 opens with it: "We're testing Sponsored Agents, which let people start a conversation with a business-sponsored agent after clicking an ad in ChatGPT." If the seller of that table is one of the advertisers in the test, it can put an agent behind its ad, and the shopper can ask all three questions without leaving ChatGPT.

That was one of several announcements in the same post. Three of the others are for the people who build the ads: an Ads Manager plugin that runs campaigns from a prompt inside ChatGPT, suggested copy and imagery drafted from your landing page, and an opt-in setting that rewrites and translates your headlines to fit each conversation. OpenAI frames the package as "reimagining both sides of advertising: how people engage with businesses in an AI-native environment and how businesses use AI to create and manage their ads."

Put plainly, the platform now offers to write much of the ad for you.

We do not run ChatGPT Ads, so nothing below is first-hand. It comes from OpenAI's post (read through the Wayback Machine, because openai.com returns an error to our fetcher), OpenAI's advertiser emails as Search Engine Roundtable published them, OpenAI's Help Center, and dated trade coverage. For the basics, how to run ChatGPT Ads covers setup and the ad unit, and conversion bidding covers what you pay for.

Sponsored Agents turn the ad into a conversation

The headline launch is also the smallest in reach. It changes what a ChatGPT ad is, but only for select advertisers so far.

What happens after someone engages with the ad

OpenAI's post describes the flow this way: "After seeing a relevant ad, a user can choose to start a clearly labeled conversation with a business-sponsored agent in ChatGPT." From there, "The user can explain what matters to them, ask follow-up questions, and follow a link to the business's website when they're ready to take the next step."

The sequence, as OpenAI lays it out:

1. A relevant ad appears in ChatGPT.
2. The user chooses to start a labeled conversation with the advertiser's agent.
3. The user explains what they need and asks follow-up questions.
4. When ready, the user follows a link to the business's website.

Look closely at step two, because OpenAI words the entry point three ways:

• The top of the post: the conversation starts "after clicking an ad".
• The body of the post: after seeing a relevant ad, the user "can choose" to start one.
• The advertiser email (Search Engine Roundtable, September 16): after clicking an ad, "users may be able to start a clearly labeled conversation with a business-sponsored agent".

Our reading: the agent is an option the user picks, not something the ad opens on its own. Plan for a conversation some users start, not one every click produces.

Three walls OpenAI draws around the agent

OpenAI is explicit about separation: "The conversation with a Sponsored Agent is distinct from ChatGPT's independent answers and separate from the original conversation that the user started in ChatGPT." Take that apart and there are three boundaries:

• Labeled. OpenAI's words are "clearly labeled conversation".
• Apart from ChatGPT's answers. The agent speaks for the business, not as ChatGPT's independent answer.
• Apart from the original chat. It runs as a separate conversation from the one the user started.

OpenAI also says, of the whole launch, "Throughout this work, our ads principles remain unchanged, and protecting the trust people place in ChatGPT remains our North Star."

The Help Center adds one line that matters to anyone putting an agent in front of customers. Its "Ads in ChatGPT" article, as archived on September 19, 2026, says: "If you choose to message an advertiser through an ad (for example, to ask questions while making a purchase decision), the advertiser will only see the messages you send them directly." That answer speaks about messaging an advertiser through an ad in general terms. Treat it as OpenAI's general rule, not as the specification of the Sponsored Agents test.

Who can run one today. OpenAI's availability line is one sentence: "Sponsored Agents are now being tested with select advertisers in the United States." Search Engine Roundtable's copy of the advertiser email repeats it: "We are testing Sponsored Agents with select advertisers in the United States."

OpenAI's Help Center also carries an article titled "Sponsored Agents in ChatGPT Ads". It returned an error to our fetcher, so we have its title and not its text. Whatever it says about eligibility, markets or pricing is not reflected here.

For planning, treat Sponsored Agents as a format you watch rather than one you budget for, unless OpenAI has invited you into the test.

Ads Manager now drafts copy and imagery from your landing page

This is the launch that goes straight at the creative.

What the suggestions are built from

OpenAI's post: "We're also rolling out AI assistance directly into Ads Manager to help advertisers create ads. When drafting an ad, advertisers can now receive suggested copy and imagery based on their landing page and campaign objective." The control sits with you: "Advertisers remain in control: they can review and edit suggestions before choosing whether to add them to a campaign."

In practice the loop has three parts:

• Input: your landing page and the campaign objective you picked.
• Output: suggested copy and imagery for the ad you are drafting.
• Your step: review, edit, then decide whether it goes into the campaign.

What the suggestions fill is a small unit, a title, a short body and one picture, laid out in detail in our ChatGPT Ads walkthrough. OpenAI's API quickstart describes the creative asset as an image: "Upload a remote image and store the returned file_id. You'll use it when you create the ad." Copy and imagery is also everything the sources for this post describe the new tool producing.

Rolling out in the post, in testing in the email. Here the two OpenAI documents of September 16 part ways. The post says "rolling out". The email to advertisers, as published by Search Engine Roundtable, says of the same feature: "This feature is currently in testing, with broader availability expected in the coming weeks." PPC Land, covering both on September 17, pointed out that they carry the same date.

Two practical consequences follow, and both are our reading:

• A missing feature is not a rejection. If you open Ads Manager and see no suggestions, your account has most likely not received the tool yet.
• Do not plan production around it yet. A tool in testing can change what it produces before it reaches you, so keep your current creative pipeline running.

Your landing page just became the creative brief

The landing page already mattered for delivery. OpenAI's Help Center says the ad system considers "the context and intent of your current conversation, the ad's landing page, title, copy, advertiser-provided context hints and targeting selections" when choosing an ad. Now the same page is also the raw material for the suggested copy and imagery.

Our view: on ChatGPT the landing page has become the creative brief, and a thin one now costs you twice, once in weaker suggestions and once in a weaker relevance signal.

Before you ask Ads Manager for a suggestion, check the page:

• One product, named plainly, with the claim you want the ad to make stated in words.
• Specs and prices that match what the ad will say, so a suggestion cannot pull a stale number.
• Nothing on the page you would not put in an ad. OpenAI describes the suggestions as "based on" the landing page. Assume any part of it can shape what comes back.
• An objective chosen on purpose, because it is the second input.

The Ads Manager plugin runs campaigns from a ChatGPT prompt

The second tool moves campaign work out of the Ads Manager interface and into a chat.

From a website or a brief to a campaign

OpenAI's post: "Advertisers can now use simple, natural-language prompts to create, update, and analyze campaigns directly in ChatGPT with the Ads Manager plugin. They can turn a website or brief into a campaign, understand performance, and get recommendations on what to do next."

So a prompt in ChatGPT can now do four jobs that used to mean opening Ads Manager:

• Build a campaign from a website or a brief.
• Update a campaign that is already running.
• Explain how it is performing.
• Recommend what to do next.

The same post makes the pitch in one line: "We are making it easier to create ads by simply writing a few prompts in ChatGPT Work." OpenAI's case for it is time, since "Many of the tasks involved in building and managing campaigns, from translating a brief into an ad to analyzing performance, can take significant time."

A preview and a confirmation before anything changes

The plugin was not new on September 16. An earlier OpenAI email to advertisers, published by Search Engine Roundtable on September 4, 2026, lists what it does in plain language:

• turn a website or brief into ready-to-launch ads
• create variants
• update campaigns
• troubleshoot delivery
• review performance
• get prioritized recommendations

The same email adds the safeguard: "The plugin previews recommended changes and asks for confirmation before applying them." Setup is three steps in its words: "add ChatGPT Ads Manager from Plugins, connect your account, and start a conversation." PPC Land's September 17 coverage reports the same confirmation step.

Our reading: the confirmation is the control that matters. Before you approve a preview, check three things:

• Scope. Which campaigns, ad groups and ads the change touches.
• Money. Whether a budget or bid moves, and by how much.
• Words and images. Whether any copy or creative changes, and whether you would have written it.

Read it the way you would read a change request from a new junior buyer, because that is roughly what it is. The words "create variants" also matter to anyone who tests creative. A variant generator is only as good as what you hand it, and creating ad variations works best when each version changes one thing on purpose.

Built inside OpenAI's own surfaces

PPC Land set the plugin beside earlier moves by Meta and X, which opened their ad systems through endpoints any compliant agent can call, whereas "OpenAI's plugin lives inside OpenAI's own surfaces". For an agency that wants one agent across several platforms, PPC Land calls that difference operational rather than philosophical. Google went its own way with agentic tools inside Google Ads.

The launch also put ChatGPT Ads inside two outside tools. OpenAI names HubSpot its first CRM partner: "Starting today, businesses managing their customers in HubSpot can connect a ChatGPT Ads account" and create ads, track performance and follow up on leads there, which HubSpot's own release calls "the first CRM integration with ChatGPT Ads". Shopify, the first ecommerce partner, got an app for US merchants that OpenAI said would open internationally from September 23, covered in our Shopify app post.

Opt-in text customization rewrites and translates your copy

The fourth launch is a setting, not an assistant, and it acts after you have finished writing.

What the setting changes

OpenAI's post: "We're also launching the ability to opt into AI-powered text customization. If enabled, it adapts an advertiser's existing headlines and descriptions to better fit the context of a conversation and automatically translates ad copy to a user's preferred language." The advertiser email adds: "Advertisers remain in control of whether this setting is enabled."

Broken into parts, the setting:

• stays off unless you opt in
• starts from your existing headlines and descriptions, not a blank page
• adapts them to the context of the conversation
• translates ad copy into the user's preferred language

PPC Land compared it with responsive search ads, which combine headlines and descriptions at serving time, "though OpenAI's version rewrites text rather than reordering it". In our reading, that difference is where the risk sits:

• A responsive search ad can only show lines you wrote.
• A rewritten headline is, by definition, a line that differs from the one you wrote.

If your category has claims you must word exactly, decide on this setting with whoever signs off your copy.

Translation into languages you never wrote for

OpenAI's Help Center already says that "An ad's language may differ from your conversation or interface language. Ad language selection can consider location, language settings, and conversation language." That gives three signals for which language an ad appears in:

• the user's location
• the user's language settings
• the language of the conversation itself

Automatic translation is what makes that workable for an advertiser who wrote in one language. It lands on a large map: PPC Land's count, after India went live on August 29, was "the publicly counted total to 41" markets, including the 31 European countries covered in our post on the European launch.

OpenAI describes the setting as working on headlines, descriptions and ad copy. Our practical reading: words baked into your image sit outside a text setting, so if you turn translation on, keep the headline out of the picture, or the ad may speak two languages at once.

The name search buyers already know

The label is not new. Google's AI Max for Shopping has a text customization feature that, in Google's words, "Generates ad copy for your Shopping ads so that they speak directly to shopper intent and conversational queries." We covered it in our AI Max post, and Microsoft Advertising has its own. The trade is the same everywhere: the system writes more of the copy, and your controls fence it in. On ChatGPT, the control OpenAI's email describes is whether the setting is enabled.

What the four launches still leave to you

Add them up and the split between platform and advertiser looks like this.

What OpenAI's tools now offer to do:

• suggest your copy and imagery from the landing page
• rewrite and translate your headlines per conversation
• build and change your campaign from a prompt
• in a US test, hold the sales conversation after the ad

What stays with you:

• the product and the claim you make about it
• the landing page the suggestions and the plugin start from
• the image you choose to put in front of the user

The picture is the part that doesn't bend

On OpenAI's descriptions, the words of a ChatGPT ad can now come from a suggestion, change per conversation and change per language. The picture is the element you place and it stays placed. On a unit this small it carries most of the argument, a point our ChatGPT Ads walkthrough makes about the format in general, and that holds more strongly once the text starts to move.

Test your creative against the suggestions, not instead of them

The Register's verdict on the creative tools, published September 16, was blunt: "Slop ads served in slop responses." It is a jab, and the risk behind it is real: if most advertisers accept the first suggestion, ads start to look alike, the problem we unpacked in our post on AI slop in ad creative. The answer is not to refuse the tools. It is to measure them.

A worked example. Take one product, one landing page and one objective, and build six ads: three image-and-copy combinations, crossed with text customization off and on.

Ad	Image	Copy	Customization
A	Suggested	Suggested	Off
B	Your product photo	Suggested	Off
C	Your product photo	Your own	Off
D	Suggested	Suggested	On
E	Your product photo	Suggested	On
F	Your product photo	Your own	On

Every ad points at the same page, so the only variables are the picture, the words and the setting. Read it in pairs:

• A against B isolates the image, with the copy held constant.
• B against C isolates the copy, with the image held constant.
• C against F tells you whether customization helps your own copy.
• A against D tells you whether customization helps the suggested copy.

Six ads is also a sensible ceiling for one product on a young channel, because each extra cell splits the same traffic thinner. Skip the cells your account cannot run yet, and read results with the discipline in our guide to creative testing.

What to find out before you plan a Sponsored Agent

PPC Land's September 17 coverage raised four questions about the format:

• what, if anything, the advertiser receives from the exchange
• which model runs the agent
• who writes its instructions
• how the unit is priced

It added a measurement question that any buyer will hit first: "Is the outcome the click that opens the agent, the number of turns, a link followed, or a sale on the advertiser's site?"

Those make a good list to take to your OpenAI contact, along with the Help Center article our fetcher could not read. The conversion tracking you already run is its own topic, covered in our post on automatic advanced matching. How much ad load each category carries, and so where testing is worth the effort, is in our read of the Comscore data.

How Novoads solves the image side of a ChatGPT ad

Novoads turns an uploaded product photo into ad images, the part of a ChatGPT ad where the new tools only suggest and you still choose, and turns the same product into UGC-style video ads for the channels that run video.

The six-ad test above needs several of your own product images before it can run, and that is usually where a small team stalls. With Novoads the work looks like this:

1. Upload the product photo once.
2. Generate a spread of ad images from it for the "your product photo" cells in the grid.
3. Turn the same product into vertical video ads for Meta and TikTok, where video is the unit.

Everything we found in the sources for ChatGPT describes copy and imagery, so treat the video as your asset for the other channels. Plans are on the pricing page.

The platform writes the words now. Bring the picture.

OpenAI spent September 16 moving more of the writing of a ChatGPT ad onto its own models: the suggestion, the rewrite, the translation, the campaign build and, in a US test, the conversation after the click. None of that is a reason to stop making creative. It is why the creative you bring now does more of the work. When the words bend to every conversation, the image, the product and the page are the ad you actually own.