The First 90 Days of OpenAI Ads: A DTC Playbook

Chapter 1Why OpenAI Ads Now

The first 90 days on any new ad channel are worth more than the next 270 combined. That window is when the auction is still undersold, when the model has not yet learned which categories it favors, and when your team's learning is uncompounded by anyone else's. Run those 90 days well and your unit economics on the channel stay better than your competitors' for two or three years. Run them badly, or wait, and you spend the next two or three years catching up.

This guide assumes you have already decided to be on OpenAI Ads. The argument for being on the channel lives in the pillar explainer. What follows is concrete: what to do in days 1 through 30, 31 through 60, and 61 through 90, what to avoid, and how to measure while the attribution data is still maturing.

Chapter 2Days 1 to 30: Foundation

The single goal of the first 30 days is to be present on the channel with creative the model can use. Not to scale. Not to optimize. Not to attribute. Be on, with creative that survives paraphrase, in your core categories. Everything else compounds from there.

Week one is connection and infrastructure. Three concrete checks before anything else. The week-one checks are tactical and easily missed; the rest of the 90 days runs on top of them.

Week two is creative prep. Write three positioning statements per category you intend to advertise. Each statement answers three questions: what the product is, who it is for, and why it is the right choice over the next-best option. Specificity beats emotion in this format. Vitamin C serum, 15% concentration, formulated for sensitive skin, third-party reviewed across 12,000 buyers gets surfaced. Glow, radiance, transformation does not.

Week three is launch with a learning budget. Pick two to four product categories, fund each with enough impressions to teach you something, and turn the channel on. Cost-per-click bidding in the $3 to $5 range is the current baseline. The minimum effective budget is whatever buys you enough conversational impressions to see daily variance, not a percentage of any other channel. Most DTC brands underspend at launch because the dollar amount looks small relative to Meta. That is the wrong comparison. The right comparison is whether the spend produces a daily signal you can read.

Week four is hypothesis-setting. Write down, in a shared doc, what you think will work in the next 60 days and why. Be specific. “The replacement-shopper archetype will outperform the researcher in our category because our repeat-purchase rate is high” is a useful hypothesis. “This channel will work” is not. The hypothesis doc becomes the artifact you measure against in chapter two.

Chapter 3Days 31 to 60: Learning Loop

The second 30 days build the feedback loop that compounds for the next 12 months. The goal is not to scale spend. The goal is to identify the two or three creative-and-category combinations the model is consistently surfacing, and to understand why.

Sit with the team once a week for an hour and ask ChatGPT the queries you think your buyers ask. Note which products surface, in what order, with what reasoning. This is qualitative work, and most analytics tools will not give you this signal. The hour you spend doing it manually is the most valuable hour of the week during this phase. Compare against your hypothesis doc. Where you predicted correctly, double down. Where the surface is recommending you in a way you did not expect, lean in. Where you expected to be surfaced and are not, the gap is almost always one of two things: your product data is thin in a way the model needs, or your category authority is weaker than you assumed. Both are fixable. Both take more than two weeks. Start fixing now.

Tighten the creative. The first 30 days produced creative that worked in the format. The next 30 are about which specific positioning statements are surviving paraphrase well and which are not. The model's recommendations are themselves data. If the model is rephrasing your positioning in a way that loses your differentiation, your positioning is not yet specific enough. Rewrite. Test. Repeat.

Build the brand-authority signals that compound. The work that makes you more recommendable on OpenAI Ads is the work that makes you more recommendable across every AI surface. Reviews, structured data, expert endorsements, third-party validation. Most of these moves take 30 to 90 days to compound into model behavior. The agent-visibility playbook covers the long-lead infrastructure piece in detail. Start the long-lead ones now.

Chapter 4Days 61 to 90: Scale

The final 30 days of the quarter are when you scale the categories and creatives that worked, retire the ones that did not, and set the hypothesis for the next quarter.

Increase spend on the winning categories. The right increase is enough to capture more inventory without flattening the variance you are reading from the model. A doubling of spend is often right at this stage. Tripling is sometimes too much because the auction in your category may not have inventory at that level yet. Watch for the point where additional spend stops increasing recommendation volume. That is the current inventory ceiling. Pushing past it wastes budget.

Retire what did not work. Not every category will succeed in 60 days. Some need a longer brand-authority buildout before the model starts surfacing you. The right call is to pause the spend, not to keep funding a category that has not produced learning. The brand-authority work continues regardless; the ad budget refocuses on the categories where the model is already cooperating.

Expand creative range on the winners. Test variants that change one variable at a time, not creative overhauls. Is it the specificity? The trust signals? The product configuration? The pricing positioning? The format rewards careful iteration, not creative pivots.

Chapter 5KPIs That Matter

Channel-isolated ROAS is the wrong primary metric for the first 90 days, and arguably for any AI ad surface. The conversion paths cross channels too often, and the dark window between recommendation and purchase makes last-click numbers misleading on both sides.

Read portfolio metrics instead. Blended CAC across all channels. Repeat-purchase rate. Gross profit per visitor. Revenue mix across surfaces. These absorb the noise of single-channel attribution and let you read whether adding OpenAI Ads helps the whole, regardless of what any single channel's last-click number says. If your blended CAC drops while OpenAI Ads is on, the channel is doing its job, even when the channel-isolated ROAS reads flat. The forecasting guide covers the modeling side of this in depth.

Build a post-purchase signal in week two. A single-question survey on the order-confirmation page, “How did you first hear about us?” with ChatGPT, search, social, friend, and other as options. The signal is imperfect. It is still signal. After 100 purchases you have a directional read on the channel mix that no analytics tool will give you cleanly for another year. The cost of building the survey is minutes. The value of having it by day 60 is enormous.

Chapter 6Common Failure Modes

Five anti-patterns surface in almost every brand's first 90 days. Each one is fixable once named.

Ninety days is enough time to learn a channel well enough to compound on it for the next two or three years. Run the playbook. Build the brand-authority asset that pays off across every AI surface. Resist the over-reactions in both directions. Trust the portfolio metrics over the channel-isolated ones.