In partnership with
Hang in there, Friday's close.
Every new ad channel has the same awkward phase. The reach is real, and nobody can prove it sold anything.
Seven months after OpenAI started selling them, ChatGPT ads are stuck right there.
Digiday spoke to a stack of agencies and measurement came up in almost every conversation. Accuracast ran a UK lead-gen campaign on ChatGPT alone and watched the form fills come in, while the platform still showed zero conversions two weeks later. The same agency saw ChatGPT report around 100 clicks where its own tracking found 20, and OpenAI's support team couldn't say why.
Other agencies haven't seen the click gap, but one, Web Guide Partner, sees conversions arrive 24 to 36 hours late, against a few hours on Google. Reporting carries no demographic or prompt data.
And plenty of brands never switch tracking on, because their legal teams won't sign OpenAI's conversion terms. The same worry slows custom audiences, which need customer lists uploaded. OpenAI's conversions API could close some of the gap, but enterprise clients are slow to implement it and smaller ones often lack the people.
Which is a bit funny, since most of them hand the same data to Google and Meta already, as one agency pointed out.
The money tells the same story. One exec has clients spending $10 million a month on Google and under $100,000 on ChatGPT, despite wanting to spend more. Accuracast can't move clients off £10,000 to £20,000 test budgets.
Its group CEO, Farhad Divecha, said the agency would need "solid verifiable information to know that it works because right now, the way the model works is you're paying for impressions and visibility."
OpenAI declined to comment.
The best advice came from Monica Shukla at Mile Marker: establish "what our measurement is going to look like before the test," and don't take the platform's numbers at face value. Obvious, sure. Most teams still do it the other way around.
So if you've got a ChatGPT test running, what number did you agree would make you scale it, and who agreed to it before the first impression ran?
Today’s sponsor:
One workspace where buyers, sellers, and AI close deals together
Instead of running your deals across scattered email threads and forgotten attachments, Aligned brings everything into one shared deal workspace.
Buyers get a single place to evaluate, loop in their team, and say yes. Sellers get visibility into what is actually happening when they're not in the room and AI Deal Insights that help you act accordingly, before it's too late.
Best part? Your first deal room takes only minutes to stand up.
TALK ON THE MARKET
Glenn Gabe of G-Squared Interactive, on a GEO experiment that went badly:
Reddit removed all four subreddits built for Otterly's 60-day experiment over automated upvotes and comments, and every AI citation pointing at them went too.
That's the risk with any visibility you didn't earn. It goes the day the platform notices.
Want to see case studies of AI-powered campaigns?
Try our sister newsletter:
NUMBERS MATTER
20%. The share of a brand's sales value on Amazon that arrives after the usual 7-to-30-day attribution window closes, by Amazon's own study of more than 2 million June campaigns, via Digiday.
That's Amazon's own study, and one agency told Digiday it saw no sales behind some off-site spend. Still, short windows undercount slow buyers.
DAILY PROMPT
Our Design Incrementality Tests to Prove AI Marketing ROI, so your ChatGPT test has a yes or no built in:
You are an expert marketing analytics strategist helping a CMO design incrementality tests to isolate and measure the true ROI of AI-driven marketing initiatives.
## Context
Our team is implementing AI in [SPECIFIC_WORKFLOW: e.g., email personalization, content generation, audience segmentation]. We need to prove incrementality, the actual lift caused by AI, not just correlation. We have limited operational capacity and need a lightweight test design that does not require months of setup or a large budget. I have attached [ATTACHED_FILES: e.g., campaign platform exports, warehouse query results, prior test readouts]. Read them, pull real baselines and variance from the data rather than assuming, run the power calculations in code, and show the inputs you used.
## Current Situation
- Current baseline performance: [METRIC_AND_VALUE: e.g., 3.2% email open rate, $45 CAC]
- AI implementation scope: [BRIEF_DESCRIPTION: e.g., AI-generated subject lines for 40% of send volume]
- Test duration available: [TIMEFRAME: e.g., 4 weeks, 8 weeks]
- Sample size capability: [VOLUME: e.g., 500K monthly emails, 50K monthly website visitors]
- Key success metric: [PRIMARY_KPI: e.g., conversion rate, pipeline velocity, customer lifetime value]
- Secondary metrics: [2-3_SUPPORTING_METRICS]
- Constraints: [BUSINESS_CONSTRAINTS: e.g., can't pause paid campaigns, limited holdout budget, brand consistency concerns]
## Your Task
Design a lightweight incrementality test that:
1. **Test Structure**: Specify the control/treatment split, randomization method, and holdout strategy that minimizes operational overhead while maintaining statistical rigor.
2. **Sample Size & Duration**: Calculate minimum sample size needed for 80% power at 95% confidence, using the observed variance in the attached data. Recommend test duration based on our volume and constraints, and state what happens if we run shorter.
3. **Measurement Framework**: Define exactly what we measure, when, and how. Include:
- Primary outcome metric and how it's calculated
- Secondary metrics that validate the mechanism
- Confounding variables to monitor and control for
- Attribution window (if applicable)
4. **Implementation Roadmap**: Provide a step-by-step setup plan that fits into existing workflows without creating new operational debt. Include:
- Data tagging/tracking requirements
- Stakeholder roles and handoffs
- Weekly checkpoint schedule
- Go/no-go decision criteria
5. **Analysis Plan**: Outline the statistical approach:
- How to calculate incrementality (difference-in-differences, propensity matching, or simple t-test)
- Sensitivity analysis to test robustness
- How to report results to CFO/Board (confidence intervals, not p-values)
6. **Risk Mitigation**: Identify 3-4 failure modes and how to catch them early.
## Output Format
Provide a concise, actionable test design document, not a lengthy academic paper. Use tables where helpful. Assume the reader is a busy CMO who needs to brief the team tomorrow. Append a short technical note with your power calculation code and assumptions so the analytics team can verify it. Flag any place where the attached data was too thin to support a firm recommendation.Point it at the ChatGPT channel instead of an AI workflow and it works as is. Use it before the test budget gets renewed on vibes.
CMO CORNER
Your attention should feel calming, not urgent.
Every new channel comes with a board member who read about it over the weekend. Set the question and the decision date, then leave the team to it.
We provide consulting services for CMOs on AI strategy for marketing teams through our Teamless offering.
TOOL UPGRADE
When the platform's numbers don't reconcile, you need your own. Three places to keep them:
Amplitude — best for B2B teams who want behavioral analytics across multi-step funnels, so a new channel's traffic shows its real path. Try Amplitude here →
Databox — best for marketing leaders who need every platform's metrics in one dashboard, with automated summaries for the exec team. Try Databox here →
AISQ — best for agencies consolidating client reporting and recurring work in one place, so a fixed team can serve more clients. Try AISQ here →
→ We have 300+ curated, filtered, and categorized AI tools in our directory. Browse to find the best ones that help your marketing team.
LEARN THIS
If you'd rather earn the AI answer than rent the ad next to it: AI SEO Mastery: Get CITED. Our free course on getting your brand cited inside AI answers.
Take it before you scale a ChatGPT ad budget.
→ We publish workshops, courses, guides, manuals, and more on AI marketing for free. Browse our learning resources for free.
BRAND SLOP
A century ago, a teenage Greta Garbo appeared in a promotional film for ball-bearings maker SKF. So naturally, SKF brought her back with AI.
The 99-second ad has an AI Garbo saying "One of my first films was for SKF, and now I'm back for my last," built with ByteDance, Google and Kuaishou models, Reuters reported. SKF worked with her estate and family. "I think we did it in a respectful way," said Per Nilsson, its head of communication and brand.
The Guardian's critic gave it one star: "Let's hope this dull AI Garbo is not the way of the future. Guardrails were never more badly needed."
The family signed off, and the review still landed on taste.