25 ChatGPT Ads Tips, According to AirOps’ Own Tests

1. Test one variable at a time, and log the status

We change one thing per test and log the hypothesis, the date, and the status. Here's where our tests stood on October 14, 2026:

If a result isn't significant yet, the table should say so.

2. Treat ChatGPT ads as a funded test, not a budget shift

Our position is that ChatGPT ads are an early channel with real potential. They deserve a funded test and a place in the channel mix, but it's too early to say whether they can replace Google or paid social.

ChatGPT ads are one channel inside AI Discovery, next to owned content, external content, and community. Paid works best when it matches what the buyer already reads about your brand.

3. Start with travel, then consumer finance and retail

ChatGPT is becoming a planning surface, not just a search tool. People now plan whole trips there, including flights, hotels, and itineraries. An ad can reach the buyer while the decision is still being made.

That's why we rank travel highest. Consumer financial services and retail ecommerce come next. This ranking is our point of view, not a measured result. Financial services also carry extra policy restrictions, so check the current policy first.

4. Keep B2B spend minimal: most of our reach was B2C mobile users

Most of the audience we reached was B2C mobile users. That won't reach a B2B ICP effectively when those buyers are on paid ChatGPT plans, which don't show ads. Our 20k-contact audience also didn't deliver (tip 18), and competitors and marketers exploring the channel clicked our ads and filled out forms (tip 23).

Run bounded B2C tests, and keep B2B spend small enough that a failed test costs little.

5. Write a one-line test brief before you spend

Every test we run starts from one sentence:

For [buyer + situation], test [one offer] against [one hypothesis]. Cap spend at [amount]. Review on [date]. Success means [qualified outcome + cost].

Before creating anything, set a spend cap, market, schedule, owner, and review date. If you can't fill in the blanks, the test isn't ready.

6. Write your scale rule before launch

Decide in advance what has to be true before you spend more. Our proposed rule says to scale only when all three of these hold:

• Conversion tracking and third-party measurement are mature.
• Reach in the target segment is large enough.
• Cost per outcome holds up against Google and paid social.

Third-party measurement is improving. In OpenAI's October update, DV Rockerbox reported that WeightWatchers' attributed acquisition cost ran 15.3% below its blended paid-search benchmark. WorkMagic reported that 67% of Dose's incremental purchases came from new customers. Triple Whale reported that 93% of Portland Leather's visitors from ChatGPT ads were new. These are partner-reported case studies, not AirOps results. (Source)

OpenAI ships product changes weekly, and today's pain points can close quickly. Reassess the rule each time a constraint closes, and log every setting change so you can still explain your results later.

7. Choose the objective first: we couldn't turn a click campaign into a conversion campaign

The objective and conversion goal are fixed after creation. We couldn't turn a click campaign into a conversion campaign or move an ad group to another campaign.

On Google, we launch on max clicks, move to max conversions, then tCPA or ROAS, all in one campaign. That path doesn't exist in ChatGPT.

Choose clicks for a traffic test or conversions for a tracked action. For conversions, validate and select the eligible event before you create the campaign. Plan a separate campaign whenever the objective changes. We're now testing click vs. conversion objectives side by side (tip 1).

8. Put each setting at the right level: campaign, ad group, ad

Most setup mistakes we've seen come from putting a setting at the wrong level. Here's our setup map:

Four terms get mixed up:

• Objective is the outcome the system optimizes toward: clicks or conversions.
• Bid strategy is how bids are set. Maximize clicks and maximize conversions are automated.
• Budget is the campaign's spending limit.
• Billing event is what you pay for. Check which settings are compatible with your objective.

Before creating a campaign, confirm account access, billing, supported markets, and measurement setup in OpenAI Ads Manager.

9. Split campaigns only by objective, budget, geography, or schedule

Inside a campaign, each ad group covers one customer need and one destination URL, with its own bid and hints. Ads within the group differ by persona, benefit, offer, or use case. A travel brand, for example, would keep family stays and business travel in separate ad groups.

Don't split a small budget across many campaigns. Each split slows how fast every campaign learns.

10. Budget extra ops time: bulk edits, filters, and reporting are basic

Day-to-day management takes longer than on Google. Here's what we've hit so far:

• Bulk changes are a pain. There's no easy copy and paste between campaigns or ad groups like Google Ads Editor.
• The basic UI is lacking. You can't filter to live campaigns only or sort by best CTR or most spend.
• The analytics view isn't mature. There's no week-over-week or month-over-month comparison and no total across channels.

Plan your structure carefully before launch, and keep your own reporting sheet.

11. Brief the situation, not the product

On search, a buyer types a few words. In ChatGPT, a buyer describes a situation: who it's for, what the constraint is, and what they're trying to do. The ad system matches ads to that conversation.

We write context hints as product or category + use case + qualifier. The qualifier can be a role, task, constraint, or platform:

This formula is an AirOps working framework, not a proven lift. Our hint-format test hasn't run yet (tip 1), and the examples are illustrative. As Hallenbeck put it in the webinar, the job is defining "the product, the situation, and the persona around who that's for." For hint syntax, see the context hints section of our best practices guide.

12. Remember that hints inform relevance but don't control delivery

Hints are contextual signals, not exact-match keywords or delivery guarantees. An ad shows only when three things line up: a conversation creates an opening, the ad is eligible, and it wins the auction. Two levers decide the outcome:

A strong hint with weak tracking or a starved budget will still underdeliver. Use campaign controls, not hints, for geography, platform, and audience restrictions.

13. Use about 30 distinct hints of 5–7 words per ad group

Our working notes use roughly 30 hints per ad group, each about 5–7 words. Those numbers are a starting point, not a proven ideal.

Group ads that share hints into one ad group. Keep hints accurate, distinct across ad groups, and relevant to the ads they share. Test hint formats separately from copy, and record what stays constant.

An LLM can quickly turn a keyword list into draft hints. Give it accurate product facts. Then have someone who knows the product remove invented capabilities and needs you can't serve.

14. Connect hint, ad, and landing page around one need

Each need gets one connected message, from hint to ad to page. Here's an illustrative brief:

Use a chain like this only if your product supports the workflow and the landing page backs up the claim. The headline and body should say who the product is for and what it does. If the promise breaks on the page, conversion drops.

15. Try a person's image instead of your logo (+7% CTR, reported in the webinar)

In the October webinar, Hallenbeck reported that swapping our logo image for a person's image, with targeting unchanged, raised CTR by 7%.

This is a reported webinar result, not a controlled test with a significance analysis. It doesn't show that faces always win, but it's a reason to test your own images. Use one clear image per concept, and don't reuse one image across personas. Check whether the extra clicks turn into qualified leads.

16. Test generated copy against your baseline: ours is +2.3% CTR so far

AirOps-generated ad copy has been live since September 28 on our click campaign. Its CTR is 2.3% higher than our baseline. That's not statistically significant yet, so we're still running it.

Run generated copy as a controlled test against your baseline rather than swapping it in. Review every claim and image before publishing.

17. Write one ad and one page per persona, need, or offer

One product serves several people, so each ad should speak to one persona, one need, or one offer. If you sell to a CMO, a head of SEO, and a performance marketer, that's three hint sets, three ads, and three pages with the same promise as their ads.

Minor wording changes don't add reach. Swapping "easy" for "simple" tells you nothing; testing "faster reviews" against "projects stay on schedule" does. Hallenbeck suggests three to five ads per ad group to start. Use fewer if traffic is thin.

18. Don't count on a small custom audience: our 20k list didn't deliver

We tried ABM-style targeting, plus excluding customers and employees, with a B2B email list of about 20,000 contacts. It never delivered.

Custom audiences reportedly need about 25,000 matched users at minimum, with 100,000 recommended. We haven't confirmed that with OpenAI yet, so treat it as reported. If your list is near that size, plan for it not to work.

19. Check beyond the first prompt: our ads often showed on the second

We often saw ads after a follow-up prompt, including the second message of a session, not the first. We also saw only one appearance per brand across a multi-thread session. Session-level exposure isn't transparent to marketers, and query-level reporting isn't available yet.

These are AirOps observations, not documented platform rules. Here's what we do about them:

• Review multi-turn conversations in placement checks.
• Measure at the session level.
• Don't build around repeat exposure. Make each ad and destination useful on its own.

20. Check who's advertising where your brand is missing

Hallenbeck saw paid ads from brands that were missing from the organic answer. He also saw review sites such as Capterra in relevant conversations. Your competition for attention may come from outside your category.

Look for buyer questions your product answers well but where ChatGPT rarely names you. Hammel calls that a chance to "play both defense and offense." Choose gaps tied to real customer needs.

21. Install Pixel and CAPI, then check the journey, not just the tag

OpenAI reports that advertisers using both the Pixel and the Conversions API (CAPI) saw a median 57% more measured conversions and a 28% lower median measured CPA than with the Pixel alone. That covers more than 1,000 advertisers from May 26 to June 22, 2026. It's an observational comparison, not a causal result, because advertisers who set up both may differ in other ways.

Set up tracking with your analytics or engineering owner before spending. Use the platform to judge delivery, and your CRM or commerce system to judge outcome quality. Our launch checklist:

Then check the journey, not just the tag:

• Run a test action.
• Verify receipt and deduplication in event monitoring.
• Check UTMs and lead or order quality in your own system.
• Compare attribution windows before reconciling reports.

Send upper-funnel events, such as page views, and lower-funnel events, such as leads and purchases. For the full technical setup, see the measurement section of our best practices guide.

22. Reconcile in your CRM: ChatGPT and HubSpot credited different campaigns

ChatGPT and HubSpot reported the same number of conversions but credited them to different campaigns. Don't trust a single platform's attribution.

Tag every ad with UTMs and carry the campaign details into the lead or order record. Reconciling in the CRM requires the CRM to already have the CAPI integration. When reports disagree, compare what each system counts, its attribution window, and the date range before you draw conclusions.

23. Separate competitor and marketer clicks from real prospects

Our notes flag clicks and form fills from competitors and marketers exploring the channel. Count them separately, qualify leads before judging CPL, and review leads with sales. As Hallenbeck put it, "Cheap clicks don't mean cheap customers."

24. Size the daily budget for about 50 conversions in 7 days

Our notes use this planning formula:

Minimum daily budget = expected CPA × 50 ÷ 7

At an assumed $100 per conversion, that's about $714 a day, or $5,000 for the week. That's an illustrative example, not a platform minimum or a measured AirOps result. Treat the formula as a floor, not your full budget. Fifty conversions also don't make a result statistically reliable.

We prefer oCPM with automated bidding. It counts both click-through and view-through conversions, which can lower reported cost per conversion compared with click-only counting. Set billing on purpose and record it when comparing campaigns. Wait for about 50 conversions before judging or rotating an ad.

25. Run a four-week validate, read, test, decide cadence, and score the result

Set the review cadence before launch. A calendar keeps the test moving, but it doesn't make a small sample conclusive. Hallenbeck recommends four weeks as the default test span:

At week 4, score the test:

Track CTR (clicks ÷ impressions) and CPA (spend ÷ attributed conversions) in the platform, plus cost per qualified lead or customer in your own system. A higher CTR alone is not a scale signal.

If the score is weak, fix the specific problem first:

• Low delivery: check eligibility, bids, budget, and audience size.
• Clicks without conversions: check tracking and the landing page.
• Poor leads: check audience fit and competitor clicks.

In the clip below, Matt and Alex cover where to find your first audience, how much budget to risk, and why each test needs at least four weeks.

Check out AirOps' ChatGPT ads platform: book a call with our team.