What to Fix After Your ChatGPT Ads Go Live

Last verified 9 October 2026. ChatGPT Ads Manager is still in beta, so reporting, bidding and optimization options may change after this date. Campaign data covers 30 September to 9 October 2026.

OpenAI says ChatGPT ads has no performance benchmarks yet across advertisers, industries or campaign types. With no average CTR or cost per click to compare against, the first weeks of a campaign come down to reading your own numbers correctly and knowing which setting to change when something looks off.

Part one covered getting a campaign live, from the first ad preview to the pixel and context hints. Once a campaign is serving, the questions change to why an ad wins or loses the auction, when the reports can be trusted, and which setting to change when a number looks wrong. The rules quoted below come from OpenAI's help center, and the numbers come from the two campaigns we ran for Search Agency, including where they did well and where they fell short.

How ChatGPT decides which ad to show

ChatGPT picks between eligible ads with a relevance-weighted, second-price auction. Your bid is one input, and how well your ad and landing page fit the conversation is the other. A closely matched ad can win against a higher bid, and because it's a second-price auction, the winner pays a price set by the competition instead of its full maximum bid. That changes where you should put your effort. In keyword-based search ads you can often buy your way to the top of a results page, but here a weak match costs you twice, since it lowers your chance of winning and the ad is less likely to be clicked even when it does show. The levers that improve relevance are the ones you control in Ads Manager, namely your title and copy, the landing page, and the context hints on each ad group.

The diagram also shows what flows back to you. Advertisers get performance reporting such as impressions, clicks, spend and conversions, and OpenAI states that advertisers never receive a user's chats, memories, name, email, precise location or IP address. You can't see the conversations your ad appeared in, so every optimization decision has to be made from the reporting columns and from what happens on your own website.

What our first two campaigns delivered

Our two Search Agency campaigns went live on 30 September and served almost all of their impressions on 1 October. Both have been paused since 2 October. Across that short run they spent Rp 813,413 for 5,873 impressions and 171 clicks, an overall click-through rate of 2.91% at an average of Rp 4,757 per click. That is well under one US dollar per click, while OpenAI suggests a starting bid of $3 to $5 per click for Clicks campaigns in its basics guide, so the Indonesian auction was far cheaper than that guidance implies.

The two campaigns behaved very differently, and the gap shows up clearly once you split the results by ad. The table lists each ad with the campaign it ran in, and the spend figures are rounded to the nearest rupiah.

The bootcamp campaign won on every efficiency measure. Its CTR was 3.69% against 1.95% for the AI Search campaign, and its clicks cost Rp 3,183 on average against Rp 8,459, which made them about 2.7 times cheaper. We can't point to one cause, because the two campaigns differed in several ways at once. The bootcamp campaign targeted all of Indonesia, promoted a training program, and used the ad and context hints Ads Manager wrote from the landing page. The AI Search campaign targeted Jakarta only, promoted an agency service, used our own keyword-style context hints, and carried the 77-character description that Ads Manager warned would be truncated. Any of those could have lowered relevance in the auction, and the next test needs to change one of them at a time.

The most useful finding came from inside the bootcamp campaign. It ran two ads in the same ad group, the speaker-photo ad from part one and a second ad created from an Ads Manager suggestion, with a dark "SEO Fighter Bootcamp 2026" image. The suggested ad received 2,356 impressions against 900 for the photo ad, and its CTR was 4.24% against 2.22%. ChatGPT moved most of the delivery to the ad people clicked more often, which is the behavior OpenAI's advice on running several creative variations depends on.

Conversions are the weak spot. The measurement pixel and the registration event were installed on the bootcamp page, and Ads Manager still recorded zero conversions from 171 clicks. Bootcamp registration is a big commitment to make on a first visit, and 120 clicks to that page over about two days is a small sample, so the result says more about how short the test was than about the channel. The context hints column also had a surprise, since the generated campaign's hints still began with "People in the United States who work in SEO", copied from the audience text in part one, and the ads delivered in Indonesia anyway. That matches OpenAI's description of hints as guidance for matching rather than targeting rules.

When to read your first numbers

Different metrics arrive at different speeds, and reading them too early leads to bad decisions. According to OpenAI's reporting documentation, impressions, clicks and CTR refresh about every 15 minutes, spend is delayed by roughly seven to eight hours, and conversions can take 24 to 48 hours to appear. A campaign that shows clicks and zero conversions on its first afternoon may simply be waiting for data to land.

Delivery also needs time to settle. OpenAI's troubleshooting guide asks advertisers to wait 24 hours after launch before reporting a campaign with no impressions, and its FAQ warns that ongoing changes to delivery systems during the beta can make delivery and spend fluctuate. We'd give a new campaign at least a full week before changing bids or budgets, because a daily budget is treated as an average across the week and one quiet day is normal.

Once there's data, use the Segment menu to break results down by device, country and platform, and export to CSV when you want daily numbers in a spreadsheet. The platform split is useful early on, since part one showed that every campaign runs on Android app, Android web, desktop web, iOS app and iOS web by default, and one of those may perform very differently from the rest.

A quick way to diagnose a campaign

Most problems show up as one of six patterns in the reporting table. The table below maps each one to the first place to look, and the sections after it explain the fixes in more detail.

When the ad isn't serving or the budget won't spend

A Not serving status always comes with a reason, which you can see by hovering over the ad in Ads Manager. OpenAI's troubleshooting guide lists the usual causes as ad copy or landing pages that don't match the ad policies, and landing pages that block OpenAI's user agents or IP addresses. The second one catches a lot of sites whose firewall or CDN treats crawlers as bots, and the fix is the same OAI-AdsBot allowance covered in part one. If an ad is rejected, you can edit and resubmit it, or request another review if you think the rejection was a mistake.

When the ad is serving but getting no impressions, work through the account before blaming the ad. Check that account verification and billing are complete, that the campaign still has daily and weekly budget left, and that the end date hasn't passed. For conversion campaigns, OpenAI also points to the conversion setup and bid amounts as common blockers.

Underspending is a separate issue. OpenAI's budget pacing spreads spend over time and doesn't guarantee a campaign will use its full budget, delivery can slow down when a campaign is running ahead of schedule, and low bids limit delivery whatever the budget. OpenAI recommends setting an end date so pacing matches the period you actually want to run, and a campaign-total budget without an end date is paced over 60 days by default. Very narrow targeting has the same effect, and in our test the Jakarta-only campaign also paid the highest cost per click, so widening it to all of Indonesia is a reasonable test.

When people see the ad but don't click

A low CTR usually means the ad doesn't speak to what the person was asking about, and the fix OpenAI recommends is more ads with more angles. Its creative guidance calls for a high volume and diverse set of ads, with multiple title and copy variations per offering, and says each variation should introduce a different angle instead of repeating the same message. The reason is the matching system. A single ad fits a narrow set of conversations, while four ads written for four different needs give ChatGPT more chances to find a relevant match.

Length matters as well. OpenAI's bulk upload checklist recommends titles of 16 to 24 characters (50 at most) and copy of 32 to 48 characters (100 at most). Our first AI Search ad in part one used a 77-character description, which is why Ads Manager warned that it could be truncated, and the preview cut it off mid-word.

To show what different angles look like in practice, the single AI Search ad could be split into four, each one written for a different reason someone might be asking ChatGPT about AI search. Our own results above show why this matters, since the suggested bootcamp ad nearly doubled the CTR of the photo ad running beside it. The four AI Search angles below haven't been tested yet.

• An outcome angle, with the title "Get Cited by ChatGPT" and the copy "Make AI assistants recommend your brand."
• A diagnostic angle, with the title "AI Search Audit" and the copy "See how ChatGPT describes your company today."
• A combined-service angle, with the title "SEO Plus AI Search" and the copy "One team for Google rankings and AI answers."
• A local angle, with the title "AI Search in Jakarta" and the copy "Indonesian team for AI search visibility."

Every one of those fits inside the recommended lengths, and each one could send people to the page that matches its angle instead of the same landing page for all four. Keep the image consistent with the message too, since OpenAI advises clear, simple visuals that match the title and copy.

When clicks don't turn into conversions

Before you change the ad or the landing page, make sure conversions are actually being counted. Start with the reporting settings. Conversions in Ads Manager combine click-through and view-through results, and you can add a separate view-through column through Edit columns. The reporting window can be set to 7, 14 or 30 days after a click, with view-through either off or at one day, and a mismatch between these settings and what you expect is a common reason numbers look wrong.

Then check the quality of the events themselves. OpenAI scores event quality over seven complete calendar days, refreshed daily, and the recommendations are practical. If you send the same conversion from both the pixel and the Conversions API, use the same pixel ID, event name and event ID so the two copies are recognized as one action, and use a different event ID for each distinct action. Server events should go out promptly after the business action is confirmed and include the real client IP and user agent, and including email or phone details where you're allowed to improves matching.

Our bootcamp campaign is a live example of this pattern, with 120 clicks, the pixel installed and zero registrations recorded. If tracking checks out and conversions are still low, the landing page is the next suspect. Someone who clicks from a ChatGPT answer arrives with a specific question in mind, so a generic homepage or a form buried below the fold loses them fast. Test the full path yourself on mobile, including redirects and the thank-you page where the conversion event fires, because OpenAI lists broken tracking paths and redirects among the things to check.

When conversions cost too much

Once conversions are tracked reliably, a high cost per conversion usually comes down to who sees the ad and how much you pay for them. Custom audiences are the most direct control. You can build them from customer or prospect lists using emails, phone numbers, their hashed versions or Google Advertising IDs, and matching takes about 20 to 30 minutes. Exclusions are the easiest win for most Indonesian advertisers, because an exclusion audience has no minimum size. Uploading your existing customers or recent buyers as an exclusion stops you paying to reach people who already converted. Inclusion audiences and bid adjustments need at least 25,000 matched users, and OpenAI recommends 100,000 or more for inclusion targeting, which rules them out for many smaller lists. When you do qualify, bid multipliers from 0.1x to 10x let you bid more for high-value audiences and less for weaker ones.

Two other levers deserve a test, one at a time. The first is context hints, where rewriting a vague list into specific phrases about who you help and when can move your ad toward conversations with clearer buying intent. The second is bidding, since Maximize results goes for volume without guaranteeing a cost per result, and switching an ad group to a manual max bid puts a ceiling on what you pay at the cost of some delivery.

Running more campaigns with bulk uploads, team access and partners

Building campaigns one screen at a time works for a first test but gets slow quickly. Ads Manager supports bulk uploads through a spreadsheet with three tabs, one each for campaigns, ad groups and ads, with up to 5,000 of each. The checklist is strict about format. Column headers can't be renamed, dates use YYYY-MM-DD, names must match exactly across tabs, there can be no merged cells, context hints go in as complete JSON lists, and image URLs must open directly in a browser from a supported host such as Google Drive, S3 or CloudFront. The bulk format currently lists Views and Clicks as campaign objectives, so conversion campaigns may still need to be set up in the interface.

Access control matters as soon as more than one person touches the account, and it matters even more for agencies. OpenAI's identity and access page describes Ad Account Admin, Member and Viewer roles, assigned from the Users & groups section, with at least one admin required on each account. Give clients or finance teams Viewer access instead of sharing a login, and remove people when they leave the project. Keep in mind that once a user belongs to 10 or more ad accounts, they can't create another one themselves.

For measurement beyond the pixel, OpenAI currently supports measurement partner integrations with Fospha, Hightouch, LiveRamp, Triple Whale and WorkMagic, and app advertisers can connect mobile measurement partners AppsFlyer, Adjust, Branch and Singular using the pixel ID and a Conversions API key. There's also a HubSpot setup guide for teams that track leads there. Most Indonesian businesses starting out won't need any of these, but an app or an e-commerce brand already using one of these tools can connect it instead of building a separate Conversions API setup.

Where Sponsored Agents fit in

The format to watch is Sponsored Agents, which lets people chat with an AI representative of a business directly inside ChatGPT after finding it through an ad. It's in a limited alpha with selected advertisers, and OpenAI says it isn't accepting early access requests yet. For now, the standard ad cards are the only way in, and the context hints, product details and landing pages you build for them are the same material an agent would need later.

Meanwhile, the work that moves results is the unglamorous part, which means several ad angles instead of one, conversion tracking you trust, exclusions for people who already bought, and a week of data before each change. If you'd like help setting that up or reading your first month of reporting, our paid search team can review your ChatGPT ads account alongside your Google Ads and tell you where the budget is best spent.