Running ads on ChatGPT sits somewhere between paid search, display, and sponsored placements — and it behaves like none of them. By 2026, ChatGPT is handling billions of weekly intent-rich conversations, which is the largest new advertising surface since mobile. If you are going to spend on it, you need a model of how it actually works first.
This is the practical guide we use with clients who are running their first meaningful budget on ChatGPT Ads — what the inventory looks like, how the auction is shaped, and where first-time campaigns quietly burn through cash.
What ChatGPT Ads actually are
ChatGPT Ads are sponsored placements embedded inside a conversational experience. Unlike search ads that live above results or display ads that live alongside content, ChatGPT Ads are part of the answer itself — a product recommendation, a sponsored action, a branded suggestion tied to what the user just asked. That changes everything about how you write them and how you measure them.
- Inline sponsored answers — a branded recommendation inside a generated response.
- Sponsored actions and handoffs — an ad that turns into a real workflow (book, buy, sign up).
- Branded plugins and tools — first-party agents the model can invoke during relevant prompts.
- Retrieval-grounded product mentions — your catalogue made available for grounded, cite-able recommendations.
How the auction is different from Google
Google auctions are keyword-bid against a clean query. ChatGPT auctions are intent-bid against a messy conversation — the user might have prompted three times before your ad becomes eligible. The bid model rewards specificity, relevance, and — critically — whether your ad actually answers the user’s question. A cheap, high-relevance ad will beat an expensive, generic one more often than it does on Google.
You are not buying a keyword; you are buying an intent cluster. Build your campaigns around the job a user is trying to get done, not the phrase they might type.
The four ad formats worth understanding
1. Inline recommendation
Your product is mentioned as part of the generated answer, marked as sponsored, with a short rationale. Highest intent capture, shortest creative surface — one line of copy often does more work than an entire landing page.
2. Action ad
The user prompts "book me a hotel in Lisbon for next weekend" — your ad surfaces a real, completable booking. Performance here is measured by action completion, not click-through.
3. Sponsored agent
A branded agent (travel, coding, finance) the model can delegate to when the user’s prompt matches. This is the most valuable format for brands with complex product surfaces.
4. Catalogue retrieval
You expose product data to the model via a retrieval feed; the model cites your products when relevant. No keyword, no bid — just structured product data and quality score.
Targeting without keywords
Targeting on ChatGPT Ads is a mix of declared context (location, language, user segment), behavioural signals (session patterns), and semantic match against the intent expressed in the conversation. Good campaigns lean more on semantic match than on demographic overlays — the model is better at parsing intent than a marketer is at guessing it.
Creative rules that separate winning ads from wasted spend
- Write like you are being quoted, because you will be. The model paraphrases everything.
- Lead with the specific job you do, not the category you are in.
- Ship structured data — features, prices, eligibility, availability — not just copy.
- Give the model one clear, completable action for every ad.
- Maintain a living feed. Stale data is penalised harder than on any other channel.
Mistakes that burn your first $10k
- Lifting Google Ads copy verbatim. The auction model cannot reward it.
- No structured product feed. You leave catalogue retrieval revenue on the table.
- Optimising for clicks instead of completed actions. ChatGPT Ads are conversion-native, not click-native.
- Treating it as brand spend. The engine is performance-native; measure like performance.
- No conversation-level reporting. If your report is query-level, you are blind.
A realistic first 90 days
- Weeks 1–3: Build structured product feeds and a testing sandbox; define 5 intent clusters.
- Weeks 3–6: Launch inline recommendation ads; measure completed actions, not clicks.
- Weeks 6–10: Expand to action ads; wire CRM for attribution.
- Weeks 10–13: Pilot a sponsored agent for your highest-revenue intent cluster.