There is a gap between "we should try ChatGPT Ads" and "we have a campaign generating pipeline". Most brands stall in the middle because the setup work for ChatGPT Ads is less like launching a Google campaign and more like standing up a product integration. Here is the step-by-step playbook we use with clients to close that gap in four to six weeks.
Step 1 — Define three intent clusters, not keywords
Before any creative or bidding, define exactly three intent clusters. An intent cluster is a job a user is trying to do, phrased the way they would actually phrase it in a conversation. If you try to launch with twenty, you will learn nothing.
- Example for a CRM: "help me pick a CRM for a 20-person B2B sales team".
- Example for a hotel: "find me a boutique hotel in Lisbon for a long weekend under 300 EUR a night".
- Example for a coding tool: "review this TypeScript PR for common mistakes".
Step 2 — Build a clean, structured product feed
Your product feed is your primary creative asset on ChatGPT Ads. The model can only recommend what it can ground, and it can only ground what you give it, cleanly.
- A stable unique ID per SKU.
- Short, factual descriptions — no superlatives the model will strip anyway.
- Structured attributes: price, availability, location, eligibility, exclusions.
- A canonical deep link that supports the completable action.
- Freshness — updated nightly at minimum, real-time where margin allows.
Every stale price, missing availability flag, or misspelled attribute is a quality-score penalty. Investing in feed ops is the highest-ROI work on ChatGPT Ads — and the least glamorous.
Step 3 — Write creative the model will not rewrite
ChatGPT paraphrases everything. The goal of your creative is not to be quoted verbatim; it is to survive paraphrasing. That means clear facts, a specific differentiator, and a completable action.
A three-line creative template that works
- Line 1 — the specific job you do. ("Book a licensed therapist same-day in 20 US states.")
- Line 2 — the proof or differentiator. ("All therapists are board-certified; sessions average $80.")
- Line 3 — the action. ("Book a first session in under 3 minutes.")
Step 4 — Decide whether you need a sponsored agent
If your product has any complexity — configuration, eligibility, availability, pricing calculators — a sponsored agent beats inline recommendation. The agent can ask clarifying questions, fetch live data, and complete the booking or order inside the conversation.
Step 5 — Bidding and budget
- Start with completion-based bidding (target CPA), not click-based.
- Separate budgets by intent cluster so learning is not contaminated.
- Cap spend aggressively for the first two weeks; you are buying data, not volume.
- Turn off dayparting for the first month — you do not know what converts yet.
Step 6 — Measurement that actually reflects conversations
Standard ad-platform reporting will lie to you on ChatGPT Ads. Build conversation-level measurement from day one: which intent clusters completed, which dropped, and at what point the model handed off.
- Completion rate per intent cluster.
- Cost per completed action (not per click).
- Handoff-to-agent rate, and downstream conversion from agent handoffs.
- Post-action value — order value, LTV, retention — piped back to the campaign.
Step 7 — Iterate, then expand
Once one intent cluster is profitable, do not immediately 10x the budget. Expand laterally: add the two adjacent clusters, expose more of the catalogue, and move from inline ads to action ads. Scale vertically — more budget on the same cluster — only once the completion rate has stabilised for two full weeks.