How to Run an Effective ChatGPT Ads Campaign

Since August 31, 2026, a European company can open an OpenAI advertising account on its own and serve ads inside ChatGPT, according to Digiday. The interface looks like Google Ads: campaign, ad group, ad. The mechanics do not. Targeting does not run on keywords, reporting never reaches query level, and a crawler validates your landing page before anything serves.

This guide walks through each step in the order you actually hit it, written for someone who has never opened Ads Manager. The placement itself, a labeled sponsored block below the answer, is set out in what ChatGPT advertising changes for brands.

Step 1: settle the objective before opening Ads Manager

A ChatGPT campaign runs on one of three buying options, set out in Search Engine Land's hands-on guide to running ChatGPT ads, published August 10, 2026: reach, clicks or conversions. The choice determines how you are billed and how delivery is optimized.

The real decision happens earlier, and it is arithmetic. The same guide reports observed costs per click between 2 and 5 dollars on US campaigns, with a 25 dollar minimum daily budget (Search Engine Land, August 2026). At that price, a campaign only makes sense if you know what a click is worth to you: your usual conversion rate on the destination page, and the value of a conversion. If you cannot answer both, you will not be able to read the result either.

One number helps calibrate the bid: Search Engine Land notes, in August 2026, that the platform surfaces an underdelivery warning as soon as a max bid drops below 3 dollars. Bidding low costs nothing, and produces nothing.

Step 2: list the conversational intents you want to cover

This is the heart of the system, and the sharpest break with classic search. In ChatGPT Ads, targeting is defined at the ad group level through context hints: plain-language phrases describing the conversations where your offer belongs. There is no keyword field anywhere in the interface.

OpenAI's help center page on creating ad groups is direct about what they are. Context hints are broad thematic signals, not exact-match rules: a hint does not force the platform to serve your ad on a given phrasing, and guarantees nothing about a specific word, topic or audience.

The documentation adds two practical instructions. A good hint is specific enough to describe a real user need, but broad enough to absorb natural variation in how people phrase things. If your hints describe meaningfully different products or use cases, split them into separate ad groups rather than stacking them into one list.

In practice, write hints the way a customer would describe their situation, not the way a vendor would name a category. "Choosing accounting software for a small team" describes a situation. "Accounting software" describes a shelf.

One honest caveat. In his comparative test, John Horn ran one ad group built from keyword-style phrases against another built from full conversational questions. The keyword-style group produced a slightly higher click-through rate and a slightly lower cost per click.

One test, one vertical, one advertiser, so do not generalize. Take the method instead: test both registers in separate ad groups rather than picking one on principle.

Step 3: write the ad inside a very tight format

The format leaves little room. OpenAI's help center page on creating ads caps the title at 50 characters with a recommended range of 16 to 24, and the body copy at 100 characters with a recommended range of 32 to 48. The cap is a technical limit. The recommended range is what displays without truncation.

You also need an image: Search Engine Land specifies, in August 2026, a PNG or JPG file, square, at least 256 pixels per side.

Three principles fit inside that space:

• A verifiable promise, not a claim. OpenAI's ad policies require the landing page to accurately represent the advertiser, product or service described in the ad.
• An action verb instead of a category noun. In 50 characters, "See what ChatGPT says about your products" works harder than "AI visibility platform".
• Several variants, not one. The help center explicitly recommends building for coverage rather than a single message: multiple title and copy combinations per offering give the system more chances to find where your ad is relevant.

On compliance, check your category before you write. OpenAI's ad policies prohibit, among others, ads advocating for or against political actors, elections, public policy or contested social issues, individual job listings, individual housing rentals or sales, and graphic or realistic violence. Other categories are restricted and require extra verification or manual review, notably financial services, regulated medical products and services, and legal services. Review itself is quick: OpenAI's developer documentation says it typically takes only a few minutes, and an ad serves only once the ad, its ad group and its campaign are all enabled and the review has passed.

Step 4: prepare the landing page and the tracking

The landing page is the step most advertisers discover too late, when an ad is rejected. Two things are decided here: OpenAI has to be able to read the page with its own crawler, and you have to be able to tie a click to a conversion. The first governs delivery, the second governs measurement.

The page must be reachable by OpenAI's crawler. OpenAI validates the destination with OAI-AdsBot, and its policies state that an ad or landing page that cannot be reviewed is not eligible to run. Its advertiser guidance asks for three checks: that the page returns a successful HTTP response to that crawler, that robots.txt allows the relevant path, and that nothing blocks automated access.

The list of things that might is long: web application firewall, CDN, bot mitigation, JavaScript challenge, CAPTCHA, authentication, geo rules. IP ranges are published at openai.com/adsbot.json for teams that filter by allowlist. An ad may need to be resubmitted once the page becomes accessible again.

Tracking runs on a click identifier. When someone clicks, OpenAI appends an oppref parameter to your URL. The measurement pixel documentation explains that the pixel reads it and stores it in a first-party cookie named __oppref, reused on later page views, and that the Conversions API can send it back with your server-side events. It is the functional equivalent of Google's gclid.

Three implementation rules follow:

1. Do not break the parameter. A redirect that strips query strings, a router that ignores them, or a landing page that bounces to a canonical URL without carrying them, and attribution is gone.
2. Deduplicate. If you send the same event through both the pixel and the Conversions API, reuse the same event ID on both sides: OpenAI processes the first event it receives for a matching key and ignores the duplicates.
3. Add your own UTMs. OpenAI's parameter serves attribution inside OpenAI. To see the traffic in your own analytics, Search Engine Land recommends adding UTM parameters to the ad URL.

One measurement subtlety is worth knowing up front: OpenAI's documentation describes a click-through window configurable at account level and a fixed one-day view-through window, with click taking precedence when both apply. View-through conversions are reported separately and do not feed standard conversion totals, cost per acquisition or billing.

Step 5: set the budget and the pace

Three settings should be decided before activation, not during: the spend cap, the bid amount, and how many campaigns run at the same time. The first two decide what the campaign can cost. The third decides whether you will be able to interpret what it produces.

The cap. Set an explicit daily budget at creation time. It is the most basic protection on this channel, and running without one is the first accident waiting to happen.

The bid. The mechanism Search Engine Land describes is a relevance-weighted second-price auction, comparable to Google's Ad Rank: your amount matters, and so does perceived relevance. Bidding under the threshold the platform flags amounts to not serving at all.

The number of live campaigns. One at a time, at least to start. The reason is structural, and it is the next step.

Step 6: read the results with the data you actually get

Ads Manager reports impressions, clicks, spend, click-through rate, average cost per click, average cost per thousand impressions and conversions, at campaign, ad group and ad level. That is the whole of what you will see, and it is enough to judge a cost per acquisition.

What it does not report matters just as much before you build a dashboard. Search Engine Land notes that no conversation data is available, nor the words or themes that triggered the ad, and that reporting cannot be segmented below country level even when targeting goes finer. There is also no public ad library for seeing competitors' ads, unlike Meta or Google: in AdExchanger, buyers name the lack of competitive data among the platform's most conspicuous gaps.

Two practical consequences:

• You test hypotheses through structure, not through the report. Since you will never know which conversation triggered the ad, the only way to isolate a hypothesis is to give it its own ad group and change one variable at a time.
• Treat market-wide figures with suspicion. Published studies on how often ads appear diverge wildly: SE Ranking found ads on 25.94% of commercial prompts across 50,006 US prompts collected on July 23, 2026 in 20 niches; Otterly.AI reported 76.4% across US shopping prompts tracked daily from July 2 to 15, 2026 in 16 industries; GrowByData put the figure at 9.2% across its client panel for June 2026, without stating which countries or how many prompts. Three methods, three definitions of a commercial prompt, three incomparable results. None of them predicts what your category will look like.

On traffic quality, one comparison is worth knowing before you commit. In that August 10 guide, John Horn compares roughly 1,500 users from Google Ads against 1,500 from ChatGPT Ads for the same US advertiser: 41 seconds of average engagement versus 17, and a conversion rate of 3.71% versus 0.21% (Search Engine Land, August 2026). One advertiser, one vertical, one period, so this is not a law. It is a reason to start small.

What changes compared to Google Ads

Six structural differences separate the two platforms, and none of them is cosmetic. They cover how you target, which audiences exist, what thresholds apply to your first-party data, how granular reporting gets, how accounts are managed, and what you can see of the competition. The table below sums them up, with sources.

Source: Search Engine Land, August 10, 2026, and OpenAI's help center.

One more difference gets overlooked: reach. OpenAI's help center states that ads may appear for users on the Free and Go plans, while Plus, Pro, Business, Enterprise and Edu accounts stay ad-free, and that no ads are shown to accounts identified as belonging to users under 18. If you sell to enterprise decision-makers, part of your audience is structurally out of paid reach. It remains reachable through the organic answer, which is what AI citation tracking is for.

Pre-launch checklist

Ten boxes to tick before the campaign goes live. They run in order through the account, the economics of a click, the targeting, the copy, the landing page, the tracking and the budget. If one of them is not green, the campaign can still ship, but you will not be able to read its result.

• The legal entity is registered in a country where Ads Manager is open, the account is verified, the favicon is in place.
• The value of a click is quantified, and the max bid follows from it.
• Context hints describe conversational situations, grouped by a single coherent need.
• Title and body sit inside the recommended ranges, not just under the caps.
• Several title and copy variants are ready for the same offering.
• The landing page answers OAI-AdsBot correctly, with robots.txt and bot protections checked.
• The oppref parameter survives every redirect down to the final page.
• The pixel and the Conversions API send the same event ID.
• The daily budget is capped before activation.
• Exactly one campaign is live, with a clear hypothesis.

If you already track your organic visibility inside ChatGPT, LightSpot's Ads tab records the sponsored cards showing on your category questions. Our agents draft a campaign in OpenAI's format, which you review before it goes, paused, into your own Ads Manager account. We wrote up how we got there in what running our own ChatGPT ads taught us.