How to Run ChatGPT Ads: Setup, the $3 Starting Bid, and the Reporting Gap

Seven markets, no keywords, and a bid floor near $3

A paid search manager opens ChatGPT Ads Manager for the first time expecting the shape of every other platform they have used: a keyword planner, an audience builder, a placement tree. What they find is a text box asking them to describe the conversations where the product might be useful, and a bid field that warns them when they go below $3. The whole setup takes about twenty minutes, and most of that is waiting for a brand review.

That is the honest state of advertising inside ChatGPT in August 2026. It is a real auction with real spend attached, and it is thinner than the platforms it is being compared to. This guide walks the setup end to end, using OpenAI's own advertiser documentation for what the platform does and clearly attributed agency reporting for what early campaigns have actually paid. If you want the wider picture of what advertising inside ChatGPT is, start there. This piece assumes you are past that and about to launch something.

What you are actually buying when you buy a ChatGPT ad

Before touching a setting, it helps to be precise about the unit. This is not a search result and not a feed placement, and the differences change what your creative has to do.

The unit is six parts, and four of them are yours

OpenAI's advertiser documentation states that "Ads appear below ChatGPT responses." Each ad is assembled from six declared parts:

• Advertiser name, from your account
• Favicon, your logo, uploaded once
• Title, the headline, yours to write
• Copy, the description, yours to write
• Landing page, the destination
• Image asset, the creative, and the only picture you choose per ad

The first two come from account setup and never change. Everything a reader actually looks at is in the last four.

Those fields are tight. In the Ads API, a creative is either a chat_card or a product_ad_template, the title accepts 3 to 50 characters, and the body accepts up to 100. That is roughly a headline and a single sentence. There is no room to explain, which puts unusual weight on the image.

Who is on the other side of it

The audience is bounded by plan and by age. OpenAI states plainly: "We do not show ads to users on Plus, Pro, or any Business plan, or to any users who tell us or we predict are under the age of 18." So the reachable population is adults on the lower-tier consumer plans, in a mode where they are actively asking questions rather than scrolling.

That matters for angle selection more than for targeting. Someone who has just asked a comparison question is further down the funnel than someone who stopped on a video, which is a different creative job from the one ecommerce video ads usually do.

The auction is relevance-weighted and second-price

OpenAI describes the mechanism directly: "We use a relevance-weighted, second-price auction to select among eligible ads." Relevance is not a tiebreaker here, it is a multiplier on your bid. And the relevance inputs are enumerated: the system "considers multiple signals, including the context and intent of the current conversation, the ad's landing page, title, copy, advertiser-provided context hints."

Read that list again. Four of the five signals are things you control, and the fifth is the conversation itself. On a platform with no keyword bidding, your copy is not just the message, it is part of how you qualify for the auction at all.

Get the account cleared before you plan the campaign

Two setup steps gate everything else, and both are easy to discover late.

Brand review is a gate, not a formality

OpenAI's partner setup documentation is blunt: "An account cannot serve ads until its brand review is approved." Reviews are described as typically taking only a few minutes, and the account exposes a review_status field you can poll. One common rejection reason is a missing favicon, which is worth fixing before you build anything.

Plan for this to happen first, not in parallel with your launch. A campaign that is built, funded and scheduled against an unapproved account simply does not deliver. The pre-flight list is short:

• Brand review approved, confirmed on the account's review_status
• A favicon uploaded and assigned, since a missing one is a named rejection reason
• Conversion tracking live if you intend to use the conversions goal
• Tracking parameters appended to every landing page URL

Ads Manager is open in seven countries

As captured on August 3, 2026, OpenAI's availability page lists Australia, Canada, Japan, Korea, New Zealand, the United Kingdom and the United States, and says the "table below shows where Ads Manager is currently available and where access is coming soon."

Do not hardcode that list into a plan. The same page states that "Country availability may continue to evolve as testing expands," and it has been evolving month to month. Check the page on the day you build the media plan.

Pick the goal, and know what it bills

A campaign takes one of three bidding types. The API reference lists impressions, clicks or conversions, and notes it "Defaults to impressions", which is worth knowing because the default is the one most advertisers do not want.

Goal	You optimize for	You pay per
Impressions (CPM)	Reach and awareness	1,000 impressions
Clicks (CPC)	Traffic	Valid click
Conversions (oCPC)	A tracked action	Valid click

The third row is the one that trips people. OpenAI states that "Billing does not change to pay per conversion. OpenAI charges you only when a valid click occurs, and the auction determines the actual CPC." The conversions goal changes which clicks the auction chases, not your billing unit. We pulled that apart in detail in how ChatGPT Ads conversion bidding works, including the prerequisites that have to be live first.

Targeting is three dials, and one of them is a paragraph

This is where the platform diverges hardest from Google Ads, and where most of the operator confusion lives.

Geography: country, region, DMA, looked up live

OpenAI's targeting documentation states that "OpenAI Ads supports country, region, and DMA targeting." A campaign that sets no location targeting can run in all available locations, which is a default worth changing deliberately.

Three levels are available, and they nest the way a media planner expects:

• Country, the default unit, and the only one that matches the reporting breakdown
• Region, for state or province level control
• DMA, for metro-level control in markets that use them

The targetable set is not a published list you can memorize. The docs point you at a lookup instead: "Use the geo lookup API when you want to search for current targetable locations." There is also a downloadable location catalog. Treat geography as a live query, not a constant.

Context hints: a description, not a keyword list

This is the replacement for keyword targeting, and it is genuinely different. OpenAI's own wording:

"At the ad group level, advertisers can provide context hints that describe the conversations, topics, or keywords where their products or services may be relevant. These hints help guide ad matching, but they are not exact-match keywords and do not guarantee delivery in specific conversations."

Three practical consequences follow:

• An ad group is now a theme, not a match type. Hints are set at the ad group level, so your account structure should split by topic rather than by keyword variant.
• Breadth is cheap. The field is documented as "Free-form audience or placement hints," and an ad group accepts up to 2,000 of them.
• You cannot prune your way to precision. Because nothing is exact-match, there is no search query report to mine and no negative to add when delivery drifts.

Custom audiences and the 25,000-user floor

First-party lists exist, with a hard size gate. OpenAI: "Note that each custom audience must include at least 25,000 matched users before it can be used. We recommend audiences of at least 100,000 users."

Two more constraints shape how you use them. "Custom audiences cannot be edited after creation," so a list refresh means a new audience and archiving the old one. And "Ads Manager does not show individual matched users or let you select specific people from an audience," which makes an audience an on-or-off control rather than something you can inspect. For most advertisers below enterprise scale, this dial is simply not available.

What to bid, and what the first campaigns actually paid

Two numbers circulate here and they are constantly confused with each other. One is a recommendation from OpenAI. The other is an observation from an advertiser. They are not the same kind of fact.

OpenAI's own starting number

The recommendation is published: "For CPC campaigns, we recommend starting with a maximum bid of $3-$5 USD per click." OpenAI also notes that "Ads Manager may also provide bid-strength guidance to indicate whether a bid is likely to be competitive or may limit delivery."

That guidance is the warning our paid search manager hit at the top of this piece. It is the platform telling you a bid is unlikely to clear a relevance-weighted auction, not a hard floor.

What one agency reports paying

For realized cost, the clearest dated figure comes from practitioner reporting. Writing in Search Engine Land on August 10, 2026, StubGroup CEO John Horn says of his agency's accounts: "We're seeing average CPCs in the $2-$5 range across industries."

Attribute that correctly when you repeat it. It is one agency's account data across its own client mix, not an OpenAI rate card and not an audited benchmark. It is useful as a sanity range for a first budget and useless as a target.

Why the bid is doing more work than it looks

Because the auction is relevance-weighted, your bid and your creative are not independent levers. A stronger ad wins impressions a weaker ad would need a higher bid to buy. The practical read: before raising a bid to fix delivery, check whether the ad is the thing losing. Raising a bid on a weak creative buys volume at the worst possible price, which is the same trap chasing click-through rate as a success metric sets on every other channel.

Reporting names the winning ad, never the winning conversation

Measurement is the part of this platform most likely to surprise someone arriving from Google Ads, and it is worth being exact rather than dramatic about it.

The metrics you get

OpenAI enumerates the set, and it is worth reading as a closed list rather than a sample:

• Impressions
• Clicks
• Spend
• Click-through rate
• Average CPC
• Average CPM
• Conversions

Conversions require tracking to be live, through the pixel or the server-side route. That pixel is also the subject of a default change advertisers should already have on their calendar, which we covered in the automatic advanced matching switch.

You can also attach your own instrumentation. OpenAI states that static tracking parameters such as UTMs "persist on ad clicks and can be used in existing analytics tools to understand traffic from ChatGPT Ads." For a first campaign, that is the highest-value ten minutes of setup available, because it moves the analysis into a tool you already trust.

One breakdown dimension, and rows down to the ad

The API's changelog records what segmentation exists: it "Added segmented insights for product, country, and device breakdowns." One optional dimension per request, across product, country or device. Rows can be aggregated down to the individual ad, so per-creative performance is genuinely available.

That last point matters and is often overstated in the other direction. You can see which ad won. What you cannot see is why.

The gap that matters

Conversation context and intent are named signals in ad selection. They appear nowhere in the reporting set. Nothing in Ads Manager or in the insights segments describes the chat that produced an impression, so the single richest input into the auction is the one input you never get back.

This is the structural fact worth designing around. On Google Ads a search terms report tells you which query bought the click, and you respond by editing the query set. Here there is no query set to edit and no report to edit it from. The feedback loop closes on the ad, not on the audience.

One advertiser's numbers, and how far they stretch

Any honest guide has to address the performance figure currently being passed around, including the part where it is being over-read.

What the comparison showed

In the same August 10 guide, Horn published GA4 data covering roughly 1,500 users who reached one website through Google Ads and through ChatGPT Ads. In that sample, Google Ads traffic converted at 3.71% and ChatGPT Ads traffic at 0.21%. Average engagement time was 41 seconds versus 17 seconds.

Why 1,500 users is not a verdict

That is one advertiser, one site, one period, and a sample small enough that a handful of orders moves the rate. It is a data point about a single account, not a measurement of the platform, and the sample size belongs in every sentence that repeats it.

The directional read that survives is softer and more useful: traffic arriving mid-conversation lands colder than traffic arriving from a typed commercial query, so landing pages tuned for high-intent search may under-serve it. That is a hypothesis to test against your own numbers, not a reason to skip the channel and not a benchmark to plan against.

So run your own comparison, and give it enough volume to mean something. A defensible first fortnight looks like this:

• Isolate the traffic. Tag every landing page URL so ChatGPT Ads sessions are separable in your own analytics, not just in Ads Manager.
• Hold the landing page constant. One destination across all ads, so the variable under test is the creative and not the page.
• Start at the recommended bid. Begin in the $3 to $5 band and change it only after the creative has been given a fair read.
• Run several distinct images, not several headlines. The image is the biggest surface in the unit and the cheapest thing to vary.
• Judge on the ad-level rows. That is the one clean signal the platform returns, so let it decide what survives.

Our guidance on how many ad creatives a test actually needs applies here more than usual, because the platform gives you fewer other things to vary, and structured creative testing is the discipline that keeps a thin-data channel honest.

How Novoads solves the creative-supply problem ChatGPT Ads creates

Add the constraints up and the operating picture is clear:

• Targeting is geography plus a description you cannot verify
• Bidding is one number, multiplied by a relevance score you never see
• Reporting names the winning ad without explaining the win

Of everything you can change, the creative is the dial with the most range, and the image is the largest piece of it, because a 50-character title and a 100-character body leave the picture carrying the argument. That is the same conclusion creative analytics keeps reaching on mature channels, arriving early here because the alternatives were never built.

That turns creative supply into the bottleneck. Testing a real spread of images means producing a real spread of images, which is exactly the step a small team runs out of budget on. Novoads takes an uploaded product photo and generates ad images through GPT Image 2 at a fixed 0.3 credits per image, so putting eight distinct treatments of one product into an ad group is a production decision rather than a shoot. The same product image can also become the UGC video ads you run on the channels where video is the unit, which keeps one asset library feeding several placements.

Start for $1. That is $1 for 3 days of access. Cancel anytime.

Creative is the only dial with real range

Every young ad platform is thin at the start, and the thin parts tell you where the leverage sits. ChatGPT Ads has taken away the keyword, hidden the conversation and capped the copy at a sentence, and left the picture and the promise entirely in your hands. On a channel that will not tell you which conversation bought the click, the fastest way to learn is to put more distinct creative into the auction and read the scoreboard it does give you. That makes production throughput the real limit, and it is worth knowing the number: across 1,045 measured renders, a finished video clip took between two and seven minutes at the median depending on the model. The mirror image is playing out on Google's side of the market, where the new Ask Advisor agents across Ads and Analytics make the measurement half cheaper without touching the creative half at all. The advertisers who treat that as the main job, rather than waiting for the targeting controls to arrive, will have a year of learning before everyone else starts.