Good morning.

Earlier this month, OpenAI opened a self-serve Ads Manager at ads.openai.com. Any US business can sign up, set budgets, bid CPC, and serve ads inside ChatGPT conversations.

The part most coverage missed: the $50,000 minimum spend that gated the original pilot is gone. Recommended starting bids sit at $3 to $5 per click. The platform now looks structurally similar to early-2010 Facebook Ads, with one difference.

The audience is 800 million weekly users in active research mode, and the competition is mostly Fortune 500 brands still figuring out whether their agency knows what to do with it.

I spent the past week reading every primary source on this rollout (OpenAI's announcement, the Digiday press call notes, the PPC Land coverage, the help docs at ads.openai.com) and stress-testing the implications with operators already running CPC campaigns on the platform.

Early performance data from a Criteo panel of 500 retailers in February showed users referred from LLM platforms converting at 1.5x the rate of other referral channels. Narrow sample, directional signal.

This is the cleanest land-grab opportunity any operator has had since TikTok Ads opened to SMBs in 2021. The playbook below gets you positioned before the operators around you catch up.

— Sam

Let’s get into it.

The headline most outlets ran was "ChatGPT Ads Manager is live."

True, but missing the operationally important parts.

Five things shipped recently:

That last one is the unlock. Owner-operators and small teams are in.

This trips up a lot of advertisers who have not seen it in the wild yet. Ads do not appear inside ChatGPT's responses. They appear below the response, clearly labeled as ads, separated from the answer. OpenAI has been religious about this separation, and every primary source confirms it. The moment users feel the answer was sponsored, the platform's value collapses.

What this means for your creative: the user has already received useful information from the model, and you are showing up at the exact moment they are looking for next steps. The ad has to read like the obvious continuation of that thought.

Free and Go-tier users in the United States, Canada, Australia, and New Zealand. Plus, Pro, and any Business plan user is excluded. Anyone identified as under 18 is excluded. So your audience skews toward people who are actively researching but have not pulled out a credit card for a paid plan yet. That is a much broader and more interesting pool than the Plus power-user crowd.

OpenAI is being deliberately conservative about advertiser categories during the beta. Currently eligible: household and consumer goods, local services, travel and entertainment, digital products, and education. Restricted categories will expand as OpenAI's review systems mature.

If you are in financial services, healthcare, crypto, supplements, or anything with regulatory complexity, you are probably not in this round. If you sell SaaS, courses, ecommerce products, local services, or travel, you are squarely in the eligible band.

Internalize this before you write a single ad: ChatGPT ads do not target keywords.

There is no exact match, broad match, or any of the search-marketing furniture you have used for the last fifteen years.

Targeting runs on what OpenAI calls "context hints," which are semantic signals from the conversation about what category the user is engaged with, what stage of decision they are in, and what kind of answer they are looking for.

OpenAI's ads lead Asad Awan gave the canonical example on the May 5 press call. A rental car company does not bid on the keyword "Yosemite." It shows up when someone is planning a trip to Yosemite, and the model derives the trip context from the entire conversation rather than any specific phrase.

The implication runs in three directions, and they reinforce each other. Bidding logic moves from keyword-level to category-level conversational intent, which makes the platform feel closer to Meta or TikTok than to Google. Creative carries more weight, because the ad itself has to earn its place when the system places it.

The Conversions API gives OpenAI's algorithm signal about what happens after the click, so the cleaner your post-click data, the better the platform gets at placing your ad in front of people who convert.

Read what Awan said in the May 5 press briefing carefully: "CPC is the way for us to make sure this is incentive compatible. We don't want advertisers to take a risk and not get ROI." That tells you exactly how OpenAI is thinking about platform health.

They want CPC to work, which means they are tuning the algorithm aggressively toward click quality. Early movers benefit from that bias.

Before you spend a dollar, run this audit. It is the highest-leverage prep work for the platform.

Open ChatGPT (the regular interface, not the API) and run three queries about your category, phrased the way an actual customer would phrase them.

For each query, screenshot the answer. Pay attention to two things: which brands ChatGPT names (these are your competitive set inside the platform), and how it describes the category (this is the language your creative needs to mirror).

You will learn three things, every time.

Your ChatGPT competitive set is different from your Google competitive set, sometimes radically. Brands that have invested in being mentioned in conversational AI (through PR, structured content, third-party reviews) often outrank the brands winning Google Ads. If you are not in that mention list, that is your first signal, and it tells you what kind of ad copy is needed to interrupt the recommendation.

ChatGPT's category framing reveals the conversational vocabulary your ad needs to use. If the model describes your category as "tools to help solo operators automate client onboarding," then ad copy that talks about "enterprise-grade workflow automation" reads as off-topic and loses the click.

The model's category description also tells you which category-level context hints to optimize toward. Build your ad groups around those category descriptions, not the keyword list you ported in from Google.

Run the same three queries against your top three competitors' categories, not just your own. Where they are being recommended, where they are not, and what gaps exist in the model's recommendations are the ad opportunities your competitors have not claimed yet. Most will not, for at least another quarter.

The biggest mistake advertisers are making in early ChatGPT ads is writing them like Google Search ads. They die. The format is different. Your ad appears after a helpful AI answer, not inside a list of competing search results. The user has already gotten information. They are at the moment of "okay, what next." Your job is to be the obvious next step.

The pattern that works in the wild follows a simple three-line structure. The first line names the specific scenario the user is already in, framed in their language and built from the question they were just asking the model. The second line delivers the single most concrete outcome you produce for that scenario, with one specific result rather than a stack of benefits. The third line offers a frictionless next step (a free trial, a demo, a specific page), never "Learn more."

The bias is toward reading like a continuation of helpfulness rather than a sales interruption. Imagine the ad as the "and here is the tool that does this" sentence the AI did not write. That is the voice.

For a course creator running paid digital education:

For a local service business:

For a SaaS or digital product:

Treat the early CTR figures as directional, not as targets. Reported numbers right now include 0.68% overall, 1.57% for top brands, and peaks at 5.4%. The market is so early that benchmarks are unstable. Your best-performing creative could be 5x your worst, and that gap is where the money is.

Here is where most operators leave money on the table. The pixel and Conversions API are live as of May 5, but they are weird in how they work, and that weirdness will cause a lot of advertisers to half-implement them.

This is the part most ad platforms do not do. You cannot generate the ChatGPT pixel yourself. OpenAI has to send it to you, based on what events you tell them you want to track. The Jellyfish team called this out on the Digiday press call: "they're taking ownership of the creation of that pixel."

Operationally, the process runs in four steps. First, sign up at ads.openai.com and complete verification. Second, specify the events you want to track (purchase, lead, signup, add-to-cart, page view). Third, OpenAI generates and delivers the pixel code. Fourth, you install it on the relevant pages. Plan for the cycle to take a few business days. Do not wait until the day you launch to request the pixel.

Per the help docs, the pixel and Conversions API support landing page views, product catalog views, page views, and add-to-cart events, plus the standard purchase, lead, and signup conversions. That is less than the full post-click picture you get from Meta or Google, but enough to run real performance optimization.

The reporting in Ads Manager Beta covers impressions, clicks, spend, CTR, average CPC, average CPM, and conversions (if measurement is configured). No individual conversation data ever flows to advertisers. Privacy is a hard line OpenAI is enforcing across the board.

Two things are explicitly "in development" with no public timeline. Cost-per-action (CPA) bidding is coming but not yet live, so you bid on clicks or impressions today and use pixel data to optimize over time. Third-party measurement is also not supported, so if your finance team requires Triple Whale or Northbeam attribution before they approve spend, you are stuck on first-party measurement until OpenAI ships the integrations.

This section is the one most "five things you need to know about ChatGPT ads" articles skip. Read it carefully.

You sign up as a US-based advertiser. Your ads are eligible to serve to Free and Go users in the United States, Canada, Australia, and New Zealand. International advertisers outside the US cannot yet sign up for self-serve directly. OpenAI recently posted job listings in Tokyo, Seoul, London, Sydney, and Sao Paulo, which signals expansion is coming, with no public timeline.

Currently eligible as of May 5: household and consumer goods, local services, travel and entertainment, digital products, and education.

Currently restricted (and likely to open over time): financial services, healthcare and pharma, crypto, adult content, anything regulated under FTC supplement and wellness rules, and most political advertising.

If you are in a restricted category, you wait. Do not try to reframe your category to fit. Verification will catch you and you will burn the account.

OpenAI manually verifies each advertiser. Glenn Gabe (G-Squared Interactive) shared his sign-up walkthrough on LinkedIn. Verification includes confirming business identity, payment method, and reviewing the advertiser category. Plan for two to five business days between sign-up and your first campaign launch.

OpenAI updated its US privacy policy on April 30, 2026 to formally disclose that it now receives purchase data from advertisers and shares data with marketing partners. If your legal team needs to review platform privacy commitments before approving spend, send them to that policy update and the Conversions API documentation. The privacy posture is strict by current ad-tech standards, but it is not zero, and the new disclosures are what make CAPI-style measurement possible.

Run this check before you launch. Every question has a real cost behind it.

A "no" on question one means you wait. A "no" on any of two through seven is a gap to close before launch. The test only generates honest signal if these answers are honest.

If $300 is a stretch right now, run a $150 micro-test over five days at $3 CPC with the same diagnostic logic. The signal is coarser. The four day-seven questions still apply.

The smallest viable test that gives you real data. Seven days, four steps, one diagnostic at the end.

$300 across seven days, about $43 per day. Enough volume to learn from at $3 to $5 CPC. Start at the recommended $3 CPC and only raise to $5 if delivery is too slow.

One Campaign, one Ad Group, three Ads. Each ad uses the same second-line outcome and the same third-line CTA. Each ad uses a different first-line scenario hook. You are testing scenario fit, not value props. Keep context-hint targeting broad in the test phase and let the platform's category routing do its job.

Request the OpenAI pixel before you launch (this alone can eat two to five business days). Install it on a single landing page that the pixel can read. Do not multivariate the landing page; you are testing creative, not pages. Pick one conversion event (free trial signup, lead form fill, or purchase). Do not track three things at once on a $300 budget.

End of week one, answer four questions with the data.

If the test worked, scale to $1,500 over the next week. Add two creative variants. Start tracking the conversational scenarios that drove your best clicks. Begin building a creative library of five to seven first-line scenarios per ad group.

If the test did not work, the diagnostic is usually one of three things: the creative did not match the conversational format, the landing page was not aligned with the conversational scenario, or your category is not where ChatGPT users are spending their decision time yet. The first two are fixable inside a week. The third is a "wait three months" answer.

The audit and the ad copy generation both run faster as prompts. Both work in ChatGPT, Claude, or any AI front end your team uses. Save them as project instructions in Claude, as a Claude Skill, or as a custom GPT inside ChatGPT so your team can run them without rebuilding the structure each time.

Paste this into ChatGPT (the regular interface, with web search enabled if you have it). Run it once before launch and then once a week while the campaign is live so you have a running record of where the model places your category.

The opening frames you as an advertiser, so the model speaks to you in commercial terms instead of giving a generic overview. The four output requirements force specificity. The "interrupt your recommendation" line produces ad-copy insight as a byproduct of the audit, which means you finish the audit with market intel and creative direction at the same time.

Use this after the audit, once you have selected three scenarios to test.

The three scenarios force you to commit to specific customer contexts before the model writes copy, which prevents the default "what does your product do" output. The "no generic benefit language" guardrail keeps the output close to operator voice. The 90-character ceiling matches the visible ad slot inside ChatGPT placements.

A practical note for teams. Assign one team member to run the Audit Prompt every Monday and drop the output into whatever channel your team uses to track market intelligence. The compound value is in watching how the model's recommendations move week over week, which tells you whether your PR, content, and review activity is making your business more or less visible inside the platform.

The Facebook Ads self-serve launch in 2010 created a window of about two and a half years where any operator with a credit card could buy ten-cent clicks and outrun every Fortune 500 brand still negotiating with its agency. By 2013 that window had closed.

The structural conditions on ChatGPT Ads in May 2026 are close to identical: self-serve access, CPC bidding, no minimum spend, 800 million weekly users with active commercial intent, a limited advertiser pool because most enterprises have not moved yet, and a platform team explicitly tuning the algorithm toward click quality so early advertisers do not get burned.

The on-ramp is about ninety minutes of work spread across two weeks. Run the Conversational Visibility Audit prompt this Saturday. Launch the $300 test next Tuesday.

The operators who do those two things now will spend the rest of the year iterating on a creative library their competitors have not built.

Talk soon,
Sam Woods
The Editor

.