Measurement is where most ChatGPT Ads programmes silently fail. The default dashboard will cheerfully report clicks, impressions, and CTR — and almost none of those numbers predict revenue on this channel. If you are going to spend real money on ChatGPT Ads, you need a measurement model designed for conversations.
Why clicks and CTR lie here
On Google, a click is an intent signal: the user actively chose your result. On ChatGPT, a click is often incidental — the user reads a recommendation inside an answer, asks a follow-up question, and the click may never happen. Meanwhile, the user who does click is often already further down the funnel than their Google equivalent. Clicks under-count the top of the funnel and over-weight the bottom.
The five KPIs that actually matter
1. Completion rate per intent cluster
How often does a qualifying conversation end in a completed action (booking, purchase, lead)? This is the single best indicator of whether your ad is doing its job.
2. Cost per completed action (CPA-C)
Not cost per click. Cost per completed action is the only like-for-like metric against your Google and paid social CPA targets.
3. Handoff rate and handoff quality
If you run a sponsored agent, what percentage of eligible conversations invoke it, and of those, what percentage reach a completion? Low handoff rate means the ad is not being triggered; low completion after handoff means the agent is the weak link.
4. Post-action value
Not every completion is equal. A first-time booking worth $40 is worth less than a customer who returns four times. Pipe order-level and LTV data back into the campaign; ChatGPT Ads reward LTV-aware bidding more than most channels.
5. Feed and quality health
- Freshness — age of the oldest item in your live catalogue.
- Coverage — percentage of your catalogue eligible to be recommended.
- Policy rejections — stale approvals are the first sign of feed drift.
Attribution — build the pipeline server-side from day one
- Every completable action emits a server-side event with a conversation ID.
- Conversation IDs are stitched to user IDs inside your CRM at authentication time.
- Attribution windows are longer than Google — 30 days is a sensible starting point.
- Do not rely on URL parameters alone; build redundancy in.
Browser-based attribution will miss a meaningful share of conversions on ChatGPT Ads because sessions are distributed across the chat surface, deep links, and apps. Server-side is the only defensible approach.
A weekly report template that actually drives decisions
- Completion rate by intent cluster (with week-over-week delta).
- CPA-C by intent cluster, compared with Google and paid social CPAs.
- Handoff rate and handoff-to-completion by agent.
- Feed health: freshness, coverage, rejection count.
- Top three winning creatives and top three losing creatives, by CPA-C.
- Incrementality check — lift on target audience vs hold-out.
Incrementality — the test that keeps you honest
Run a rolling incrementality test: hold out a defined user segment from ChatGPT Ads exposure, measure the lift in completed actions for the exposed group vs hold-out. This is the only measurement that survives a CFO review. Platforms that cannot support a hold-out should be used with caution and explicit caveats in reporting.
Reporting anti-patterns to retire
- Clicks and CTR as headline metrics — bury them or drop them.
- Last-touch attribution — stop using it as the only lens.
- Average order value as a headline — use LTV; AOV flatters short-term campaigns.
- "Engagement" metrics invented by the platform — demand completed actions.