How to Set Up ChatGPT Ads and What Creative Actually Fits

Set a fixed test budget, a single market you can actually reach, and a success definition that is not 'match our Reels CPA in week one.' Inventory is currently aimed at logged-in adults on Free and Go tiers in rolling geos, not Plus-class professionals by default. If that audience is not your buyer, stop here. If it is, proceed knowing specs, bidding, and reporting will move under you.

OpenAI has added self-serve access, product feeds, custom audiences, conversion-optimised bidding, and new markets in bursts. A setup guide written in spring will already be stale on a button. Your operating rule is: confirm the current campaign types, character limits, and conversion setup in ads.openai.com help on the day you build. Screenshot the review screen. Do not copy a colleague's structure from memory. The product is maturing; your process should assume the form, not the fields, stay stable.

Contrast the job with Meta and TikTok UGC video. Those channels buy a mute scroll with a face, a first frame, and spoken proof. ChatGPT buys a moment after someone already asked a question. The user is in a research or decision conversation. Your card has to be a useful next step, not a pattern interrupt. If your only creative machine is vertical talking-head video, you still need that machine — for those platforms. Forcing the same file into ChatGPT is how you learn the wrong lesson about 'AI ads' when you actually tested a format mismatch.

Go to ads.openai.com, sign in with the OpenAI account that should own spend, and complete business details: legal name, website, square logo or favicon, industry, country, currency, and time zone. Set up the billing profile next. Do not design ads in a vacuum while the account is still in a half-finished onboarding state that cannot serve, and do not share login workarounds instead of inviting users.

Use the brand name and mark exactly as they should appear on the card. This is account-level identity, not a per-ad logo swap like some social platforms. If the favicon is a messy app icon, the card will look like spam next to a clean answer. Have a square PNG ready. Confirm who on the team needs access and invite them after billing works, not before, so you are not debugging permissions and creative in the same hour. If conversion optimisation is available on your account, install the pixel or Conversions API now rather than after you have a week of unattributable clicks.

Check whether your market is actually in the current rollout. Availability has expanded in waves (US first, then other regions). If the UI will not let you target the country you sell in, you do not have a creative problem. You have a coverage problem. Do not 'fix' that by targeting a nearby country with a mismatched landing page. Wait for the geo or run the test where you can fulfil. Maturing products punish clever workarounds more than they punish patience, and a wrong-geo click is a wasted lesson.

Create a campaign with a single objective you can live with — clicks, reach, or conversions if that bidding type is live on your account — a budget you can hold for at least a week, and one offer. Add an ad group with a default destination URL. Resist catalog-everything on day one. Thin reporting plus many SKUs is how you learn nothing and still spend the test budget.

The structure is familiar on purpose: campaign, ad group, ad. Campaign holds budget, dates, geo, and objective. Ad group holds bid and context hints. Ad holds the chat card: title, body, image, URL. If product-feed campaigns exist on your account and you are ecommerce, consider a feed campaign for SKU matching and a separate standard campaign for a flagship offer. Do not mix a brand story and a 40-SKU dump in one ad group. You will not see which conversation matched which item, and the card will look like a catalog accident.

Name objects like a grown-up: market, offer, hint theme, date. Future you will not remember what 'test 3' meant. Set end dates on tests so a forgotten beta campaign cannot keep spending after the team moved on. If custom audiences are available, use them as include or exclude lists you understand (buyers, high-value customers, unsubscribers), not as a substitute for context hints. First-party lists refine; they do not describe the conversation. The conversation is still the targeting primitive on this platform.

In the ad group, describe the chats where your product is a sane next step: who is talking, what they are trying to do, and what you are not for. Context hints guide matching; they are not exact-match keywords. A list of product nouns without intent will show you next to the wrong answers. A paragraph of real questions will not. Write the questions first, then trim them into the hint field.

Write hints the way a user writes to ChatGPT. 'I'm comparing travel backpacks that fit under a seat and do not scream laptop bag' is a hint. 'backpack, travel bag, carry on, rucksack' is a keyword leftover. Include disqualifiers: 'not for children,' 'not medical advice,' 'not enterprise procurement' if those conversations would waste spend. Cover three layers if you have the room: the job-to-be-done, the comparison moment, and the objection ('is a $X bottle actually worth it'). Then stop. A novel is not more precise. It is more ways to match badly.

Build two ad groups only if you have two genuinely different conversation types (first-time comparison versus replenishment, or two products that should never share a card). Duplicate hint soup across ten ad groups is theatre. After launch, you still will not get a transcript of the chats you appeared in — reporting has been limited to delivery and click metrics for much of the rollout — so the hint has to be right at write time. You are targeting in the dark more than on Google Search. Write like you will never see the query log, because you may not.

Write a title within the live character cap (plan for 50, aim shorter), a body within the live cap (plan for 100), a square product image, and a landing page that continues the conversation. The card sits under an answer. It should feel like a labeled, useful option — not a creator yelling. If your first draft could be a TikTok hook, it is the wrong object.

Lead with the specific job the user is already in: 'Under-seat backpack with a real laptop sleeve' beats 'Wait, packing just changed.' The second line is a Meta hook. The first line is a chat card. Use a true constraint (size, time, what is included, who it is for). Put the offer on the landing page if it will not fit. Truncation is common; front-load the noun. The image should be a clear square pack shot or in-use crop that matches the PDP, not a 9:16 talking head with captions that become unread noise when cropped to a card thumbnail.

This is the opposite of Meta and TikTok UGC video, where a face, a pattern interrupt, and spoken proof do the work in a mute feed. ChatGPT already supplied a long answer above your card. Your job is to be the clean sponsored next click, clearly labeled, visually separated from the model text. Do not try to out-entertain the answer. Try to be the obvious product for the question that was asked. If you sell on Shopify, the landing page should open on that exact SKU with the same name you used in the title. Conversation-to-PDP mismatch is more punishing here than in social, because the user is mid-research and will notice.

Start two or three cards in the same ad group, or two hint themes if you truly have two intents. Run for a full week at the test budget. Read spend, clicks, CTR, and conversions if tracking is live. Do not expect audience breakdowns, query reports, or a hook-rate dashboard. Thin reporting is part of the product today, not a setting you missed, so keep the test small enough to judge without those cuts.

Change one thing per iteration: the hint paragraph, or the title noun, or the landing page. If you change all three, the channel is too young to tell you why CPA moved. Keep a log of UI changes you notice (new bid types, new geos, new card layout). When OpenAI ships a format refresh, pause long enough to rebuild the card rather than letting an old crop sit in a new template. Maturing ads products often change the unit faster than they change your habit, and old crops look worse after a layout refresh.

If conversion bidding is available and you have enough events, try it only after the pixel is firing on a real thank-you event, not a page view. If it is not available, do not pretend CPC is ROAS. Call the test what it is: qualified traffic from high-intent chats. Compare landing-page conversion rate against your Search traffic, not against TikTok. Different temperature of intent. A ChatGPT click that bounces is often a hint mismatch. A TikTok view that does not click is often a hook mismatch. Use the right autopsy.

Do not pause Meta and TikTok creative testing because a new AI ad unit exists. Continue to ship vertical UGC for those auctions. Use ChatGPT Ads for the research moment, and social video for the scroll. The teams that blur those jobs will declare both channels 'inconsistent' when they actually ran the same asset in two incompatible slots and measured them with the same KPI.

Klip Kanvas still belongs on the Meta and TikTok side: product URL in, talking-head UGC out, hooks tested in 9:16 with captions. That output is the wrong native file for a ChatGPT card, and that is fine. Steal language, not files. The customer nouns that win as spoken hooks can become the title nouns on the card after you cut them to length and strip the performance cadence. A confession that stops a thumb is rarely the line you want under a sourced answer. Translate the proof. Do not paste the script.

If OpenAI later adds richer media units, re-qualify the format then. Do not pre-build a vertical video library 'just in case' the chatbot becomes TikTok. Build against the unit that exists: square image, short text, conversation match. Revisit this split quarterly, because the product will change. Until a true video unit is live and documented, your creative system should have two folders with two briefs. Mixing them is how a maturing channel eats a working social program you already knew how to test.