Five months ago, OpenAI's advertising platform was a $200,000 minimum experiment for enterprise brands with patience and budget to burn. Today, any US business can log into a self-serve Ads Manager, set a CPC bid, and start running ads inside ChatGPT conversations. The question for marketing leaders isn't whether this channel exists. It's whether the math works for your pipeline.
I've spent the past several weeks pulling apart the mechanics, pricing, and measurement infrastructure. Here's what I'd tell a CMO walking into a board meeting next week: the channel is real, the targeting is fundamentally different from anything you're running today, and the measurement gap that plagued early adopters has narrowed considerably. But the unit economics require a different mental model than Google or Meta.
What Actually Changed in May 2026
OpenAI's May 5 announcement transformed ChatGPT ads from an enterprise-only pilot into a self-serve platform. The $200,000 minimum floor dropped to zero. CPC bidding arrived alongside the existing CPM model. A Conversions API and measurement pixel shipped, closing the attribution gap that made early tests impossible to evaluate.
The platform now supports both CPM and CPC bidding. According to Digiday, recommended starting CPC bids run $3 to $5 per click, with bids under $3 routinely failing to clear delivery thresholds. CPMs have compressed from the pilot's flat $60 to a range of roughly $25 to $60 depending on category. AuthorityTech reports the pilot hit a $100 million annual run rate within six weeks.
The geographic footprint expanded rapidly. PPC Hero's guide tracks the rollout: Australia and New Zealand in March 2026, US self-serve access on May 5, then Japan, South Korea, and the UK in June. The UK became the first European market to launch ChatGPT ads.
The Targeting Model Is Not What You're Used To
ChatGPT ads don't target keywords in the traditional sense. They target conversation context. When a user asks "what's the best project management tool for remote teams," relevant software ads can surface within that answer. The targeting mechanism matches ads to the topic the user is discussing, not to a keyword bid auction.
This matters for B2B marketers because the audience composition is unusual. Vision Media's analysis notes that ads appear only to users on the Free and Go tiers. Plus, Pro, Business, Enterprise, and Edu subscribers don't see ads. Your most senior prospects, the ones on plans funded by their employer, are likely on ad-free tiers. The audience that does see ads skews toward individual contributors, founders at smaller companies, students, and researchers using the free version for personal tasks.
Stackline's data shows nearly 100 million shopping questions flow through ChatGPT weekly in the US alone. For the Headphones category, ChatGPT generated 19.4 million shopping questions over the last 52 weeks, surpassing Amazon Alexa at 10.68 million. The volume is real, even if it doesn't approach traditional search scale.
The Measurement Stack: What You Get, What You Don't
The February 2026 pilot gave advertisers a weekly CSV with clicks and impressions. That was it. Top Growth Marketing's updated analysis confirms the April and May updates closed most of the measurement gap. ChatGPT Ads now ship with the OAIQ pixel for site-side event tracking, a Conversions API for server-side attribution, CPC bidding, and configurable attribution windows.
You see views, clicks, and conversions in OpenAI's Ads Manager. What you don't see, and won't, is user-level data, conversations, or demographic breakdowns. Reporting is aggregated by design. Stape's configuration guide recommends hybrid tracking (pixel plus Conversions API) to improve accuracy, especially under browser limits and fragmented journeys.
The structural constraint hasn't changed: OpenAI does not share individual user data, conversations, or behavioral profiles with advertisers. This is closer to CTV measurement than to Google Ads granularity. For B2B marketers accustomed to seeing which company visited which page, this is a significant adjustment.
The Unit Economics: Running the Numbers
The Growth Syndicate reports click-through rates run seven to ten times lower than Google Search, but conversion quality is 1.5 times higher. At a $4 CPC with a 0.5% CTR, you're paying roughly $8 per click on an effective basis. Compare that to Google Search CPCs in competitive B2B categories, which often run $15 to $50 per click, and the math starts to look interesting.
Digital Applied's benchmark data from beta advertisers in B2B SaaS, financial services, and premium retail shows conversion rates 2 to 4 times higher than equivalent Google Search campaigns. The audience skews toward high-income, highly educated users asking specific, purchase-proximate questions.
The auction uses a relevance-weighted second-price model. North Country Growth's analysis explains that your maximum bid sets a ceiling, not what you actually pay. The clearing price is determined by the next highest competing bid, adjusted for relevance. If OpenAI's system scores your ad as highly relevant to the conversation context and the competing bid is lower, you pay less than your max.
The Ads Manager: What You're Working With
PPC Hero's step-by-step guide walks through the interface. The Performance Trend dashboard shows spend, impressions, clicks, and CPC across date ranges. You can segment campaigns by device or country. The left menu includes Campaigns, Tools (change history), Conversions (data source setup, event creation, implementation), Feed (product feed upload), Billing, and Settings.
Campaign creation follows the standard hierarchy: campaign, ad group, ad. The platform supports both CPM and CPC bidding at the campaign level. June 2026 updates added the ability to edit campaign budget types, clone CPM campaigns to CPC, set custom max CPM bids, and make bulk edits directly in the interface.
Conversion tracking requires implementing OpenAI's measurement pixel or Conversions API. The attribution window defaults to seven-day click and one-day view, configurable per campaign. UTM parameters work for traffic tracking.
The Pilot Plan: Two Weeks, Three Checkpoints
For a B2B marketing team evaluating ChatGPT ads, I'd structure a two-week pilot with a $5,000 to $10,000 budget. The goal isn't to prove ROI in fourteen days. It's to establish baseline CPCs, CTRs, and conversion rates that let you model what a scaled investment would look like.
Week one: Launch three to five ad groups targeting different conversation contexts relevant to your product category. Use CPC bidding at $4 to $5 per click. Implement the Conversions API for server-side tracking of demo requests or signups. Set up UTM parameters for traffic attribution.
Week two: Evaluate delivery volume, actual CPCs (which should clear below your max bid), and conversion rates. Compare cost-per-lead to your existing Google and LinkedIn benchmarks. Model what a $25,000 monthly investment would yield at observed rates.
Checkpoint one (day 3): Confirm ads are delivering and pixel/API is firing correctly. Checkpoint two (day 7): Review CTR and CPC by ad group, pause underperformers. Checkpoint three (day 14): Calculate cost-per-conversion, compare to existing channels, decide on scale or kill.
Risks and Mitigations
The audience composition risk is real. If your buyers are enterprise decision-makers on paid ChatGPT plans, they won't see your ads. Mitigation: use ChatGPT ads for mid-funnel awareness with individual contributors who influence purchase decisions, not for direct executive targeting.
The measurement gap, while narrower, still exists. You won't get the user-level attribution you're used to from Google. Mitigation: layer in branded-search lift analysis, direct traffic spikes during flights, and post-purchase "how did you hear about us?" survey responses.
The platform is still maturing. Features are shipping monthly, pricing is volatile, and best practices are being written in real time. Mitigation: treat this as a learning investment, not a core channel. Budget accordingly.
The math can work. But only if you model it first, run a tight pilot, and hold the results to the same CAC payback standards you'd apply to any other channel. Model or it didn't happen.
