ChatGPT Ads: how to build your first campaign

ChatGPT ads are no longer a closed pilot. There is a self-serve Ads Manager, it is open in 56 countries including Israel and the United Kingdom, and you can build and launch a campaign yourself.

This guide walks the whole setup, from opening the account to reading the first numbers, and ends with a prompt that writes your context hints for you.

Three things decide whether this is worth your time at all.

One detail catches people out. Country, billing currency and time zone are fixed when the account is created and cannot be changed afterwards. Getting them wrong means opening a new account, so decide which entity is advertising before you touch the form.

Everything happens at ads.openai.com. Sign in with an OpenAI account on your work email, or create one.

Verification and account review are two separate steps. Passing verification does not always mean you can run campaigns the same day, so watch the account status rather than assuming. Do not submit a second application while one is in review. Duplicates do not jump the queue and they confuse which account is the real one.

One login can hold several advertiser accounts, which is how an agency keeps clients apart.

Three levels, and they do not behave quite like Google or Meta.

You can build through the guided flow, or bulk upload a completed campaign schema when you are running at scale.

The objective decides how you are priced and what delivery optimises towards.

Choose a campaign total budget or a daily budget. The total budget is the stricter control. The daily budget is a target rather than a cap: a single day will not exceed twice what you set, and a week will not exceed seven times it, with the week running Sunday to Saturday in the account time zone.

If the campaign burns its weekly allowance early it can stop delivering until the next budget week, even while it still shows as active. Worth knowing before you panic about a quiet Friday.

Minimum daily budgets depend on the billing currency:

What to expect on price. Clicks currently run roughly $2.50 to $8.00, against $1 to $3 on Google Search. It is more expensive per click because the audience is mid-decision rather than mid-search. Bidding is a second-price auction, so if you bid $20 and the next advertiser bids $10, you pay just above their $10 and not your maximum.

Give a campaign enough runway to gather real data before judging it. Country targeting has a restriction worth planning around: a new self-serve account can often advertise only in its own home country at first. Selecting other countries does not override that. The account becomes eligible for other markets once identity verification is approved and it has accumulated real spend at home.

Platform targeting lets you choose any of Android app, Android web, Desktop web, iOS app and iOS web, and you can segment performance by the same five.

An ad group should cover one shared product, service, theme or customer need. Split anything that genuinely needs different messaging or a different landing page.

A plumbing business is the clearest example: emergency repairs and preventive maintenance belong in separate ad groups. Same business, completely different moment for the customer.

You will find the field inside the ad group, under Ad delivery & tracking, labelled Context hints (optional). Ignore the word optional. It is the most important field on the platform, and leaving it empty hands the matching engine nothing but your title, copy and landing page to work from.

Context hints are the part with no equivalent on other platforms, and the part most people get wrong. They tell the system when your ad may be relevant, in plain language. They are not keywords, not exact match, and not a targeting rule, and they do not guarantee delivery against any particular conversation.

The difference is specificity. “Running shoes” tells the system almost nothing. “Cushioned everyday running shoes for beginners training for their first 5K” tells it who the shoe is for and when it matters.

The prompt at the end of this guide writes five of these for you.

This is where habits from keyword platforms actively hurt you. ChatGPT is not matching a query to a keyword. It reads the intent of a live conversation, alongside your landing page, title, copy and context hints. Copy that spells out practical value beats a clever line.

One technical trap that will silently kill a campaign: your landing pages must be reachable and must not block OpenAI’s crawlers, OAI-AdsBot and OAI-SearchBot. If a client’s site blocks unknown bots by default, fix that before you spend a shekel.

Campaigns go to review before they serve. Once live, check that the ads are actually serving. If one shows Not serving, hover the status for the reason. Most causes are fixable by editing and resubmitting, though some ads are rejected on policy, and delivery also stops when the account needs verification or a billing action.

Ads Manager reports impressions, clicks, spend, CTR, average CPC, average CPM, and conversions when measurement is set up. You can read it in the table, in trend charts, or as a CSV export.

Two reporting quirks to know before they mislead you. Impressions, clicks and CTR refresh roughly every fifteen minutes, but spend can lag seven to eight hours. A zero in the spend column next to live clicks does not mean nothing has been charged.

For measurement beyond the click, there is a pixel and a Conversions API. Reporting comes back aggregated. No advertiser sees individual conversations.

At two and a half to eight dollars a click, this channel cannot be judged on clicks. Set up measurement before you raise a budget, not after.

There are two routes and they are not exclusive. The JavaScript pixel goes on the site like any other tag. The Conversions API sends events server to server, which survives ad blockers and browser restrictions. Running both is the sturdier setup, exactly as you would on Meta.

Once tracking is live you unlock the Conversion optimization objective, and that is where the real decision sits. You choose whether you are billed on impressions or on clicks, and the two behave very differently.

OpenAI is unusually blunt about this one: choosing Clicks will lower your recorded conversions. That is not a bug. Someone researching a decision in ChatGPT often does not buy in the same session, and click-only attribution never sees them. If the purchase cycle is longer than a single sitting, impressions billing reflects reality better.

Two limits to know. You need at least one supported standard conversion event, and custom events are not accepted for this objective. To move an existing campaign onto it, you clone the campaign rather than switch it in place.

If the client is a shop rather than a service, this is the version to build. Feed-based ads are reported to be among the strongest performers on the platform, because the ad is assembled from structured product data rather than written by hand, so it can carry the image, title, price, sale price, rating and brand.

You will see the choice at the very top of campaign creation, in the Campaign type dropdown: Standard or Product feed.

The trap is the CSV. Feed items expire after two weeks, so a manual upload quietly empties itself mid-month. Use a hosted URL or automated SFTP for anything you intend to leave running.

Two more constraints. Feed campaigns target and exclude at country level only, with none of the finer geography a standard campaign allows. And during the beta your feed products are eligible for ads only, so they will not surface inside ordinary ChatGPT conversations.

Take an online shop selling baby and toddler equipment, the kind of account most agencies have a version of. Here is the whole build.

Objective Clicks, because the goal is traffic to product pages rather than reach. Daily budget 60 shekels, which is the Israeli minimum, deliberately small while we learn. One country. Platforms left open until the data says otherwise.

The instinct is to split by category: strollers, cots, car seats. That is the wrong cut here. The same stroller is a completely different conversation depending on who is asking, so split by the moment instead.

Three per ad group, written as sentences rather than keywords. For ad group A:

For ad group B:

Compare those with what most people submit. “Baby products” and “strollers online” describe the shop’s inventory. They tell the matching engine nothing about who is asking or why, which is the only thing it needs from you.

This is where a habit from Meta will cost you money. A promotional line lands badly inside a conversation.

The second one answers the question the person is already asking. Write three or four distinct variations per ad group, each taking a genuinely different angle, so the system has more chances to find a moment where you fit.

Ad group A goes to a comparison or buying-guide page, not the homepage and not a single product. Someone mid-research wants to keep researching. Ad group B goes to a curated gift collection with prices visible.

Tag both. Under Advanced on the campaign there is a Tracking parameters field, and it accepts placeholders that Ads Manager fills in for you: {campaign_id}, {ad_group_id}, {ad_id}, {ad_account_id} and {oppref}. Set it once at campaign level and every click arrives already labelled, which is currently the most reliable way to see which conversations produced revenue.

Not CTR, and not cost per click. At 2.50 to 8.00 dollars a click this channel only justifies itself on cost per conversion, so set the pixel up before you launch rather than after. If a campaign will not scale, the usual fix is to widen the context hints further up the funnel rather than to raise the bid.

More and more people use ChatGPT to research topics, compare products, ask for recommendations and solve problems. ChatGPT ads let a business appear while those decisions are still taking shape.

Enter your product details and you get five context hints ready to paste into Ads Manager, each with a short note on what it contributes, and a flag against anything that could not be verified.