ChatGPT Ads generated $0.89 in revenue per session versus $2.45 for paid Google Search in the US traffic measured by a new retailer study. Conversion rates were 0.37% and 1.17%. These early results support cautious, measurable tests before scaling budgets, not a universal prediction for every advertiser.
The October 2026 working paper by Maximilian Kaiser and Christian Schulze is under review. Its paid-channel comparison is observational, and its headline figures describe a particular sample and period. Here is how to use the evidence without turning it into a promise or a verdict on the entire channel.
What the study measured
The paper, Can Advertising Fund Free AI? Evidence from ChatGPT’s International Advertising Rollout, covers 60 weeks from June 29, 2025 to August 22, 2026. Different analyses use different samples:
- 1,898 websites form the organic-referral panel.
- 387 websites with paid ChatGPT traffic form the paid referral panel used for engagement analysis.
- 102 websites have transaction and revenue tracking for the financial analysis.
The financial results therefore do not come from all 1,898 websites. Advertisers chose their channels rather than being randomly assigned. The comparison describes observed performance; it does not prove that moving a business’s budget between channels would reproduce the reported gap.
Two useful numbers, with the right denominator
For US traffic in the financial sample, paid ChatGPT recorded a 0.37% conversion rate versus 1.17% for paid Google Search. Revenue per session was $0.89 versus $2.45, roughly 64% lower for paid ChatGPT.
Revenue per session is retailer revenue associated with traffic. It is not advertiser profit, OpenAI revenue, or return on ad spend. Assessing acquisition economics also requires traffic cost, margins and consistent conversion definitions.
These pooled figures are not forecasts for every category or a benchmark for Vietnam. They describe an early period of the advertising product, with a particular set of websites and users.
How to read the six-panel chart
The chart shows conversion rate, order value, revenue per session, bounce rate, pageviews per session and session duration. Each bar is an adjusted difference from direct traffic, not a raw channel conversion rate. The models account for website, device and week differences; the whiskers are 95% confidence intervals.
Conversion and bounce rate use log-odds. A negative value on those axes cannot be read as an equivalent percentage-point decline.
In these models, paid ChatGPT has lower conversion, order value and revenue per session than direct traffic, along with higher bounce rate and lower pageviews and duration. Panels A–C use the 102-site financial sample; D–F use the 387-site paid referral panel.
Keep the pooled US levels separate from these adjusted estimates. They answer different questions. In particular, pooled order values and adjusted within-website comparisons can move in different directions.
Figure 4: adjusted channel differences versus direct traffic. Panels A–C cover 102 websites with financial data; D–F cover 387 websites with paid ChatGPT traffic. Whiskers show 95% confidence intervals. Source: Maximilian Kaiser and Christian Schulze (2026), October working paper. Reproduced with permission.
Why “six times worse” needs more context
The paper calculates a parity click price: the ChatGPT CPC that would produce the same tracked revenue per advertising dollar as paid search. It combines the US revenue-per-session ratio with a $0.72 median Google Search CPC from 16 US-dominant accounts, yielding about $0.26 per click.
The authors compare this with a $1.72 ChatGPT CPC from one publicly disclosed account. The difference of more than six times comes from that calculation. It is not a directly observed average ROAS comparison across the entire ChatGPT advertiser sample.
The calculation assumes similar margins and counts the tracked visit. The paper’s resampling interval for the parity price is broad, about $0.05–$0.57, and one advertiser accounts for 76% of US paid ChatGPT revenue. These details matter when translating an aggregate result into a budget decision.
The data does not see every purchase journey
The study tracks the referring last click into a website. It does not observe every prompt, ad impression or previous brand interaction.
Someone could discover a product in ChatGPT and return later through Google or direct traffic. That possibility should be measured, not assumed to make up the performance gap. Failure to detect an indirect lift is also not proof that awareness or later-purchase effects never occur.
Paid and organic ChatGPT referrals need separate measurement. Sharing a platform does not mean the visitors have the same intent or business value.
GPTAd’s recommendation: define the decision before the test
- Choose a business outcome. Use profitable orders or qualified leads if those are the commercial goals, rather than treating clicks as the final result.
- Separate traffic sources and validate tracking. Distinguish paid ChatGPT, organic ChatGPT and paid search. Check conversion events and revenue before launch.
- Compare like with like. Align market, timing, products and conversion definitions. Record differences in offers or landing pages that could affect performance.
- Set a test budget and stopping rules. Decide what can be spent on learning and which acquisition-cost thresholds justify more spend. Avoid conclusions based on only a few purchases.
- Measure later value where relevant. For longer sales cycles, track lead quality, delayed purchases and returning customers. Use a control group to assess incremental value when feasible.
The study is a useful reference for initial expectations. A business’s budget decision still needs its own consistently measured costs, revenue and margins.
Further reading
Choosing a first ChatGPT versus Google Search test
Separating organic ChatGPT referrals from paid ads
References
Kaiser and Schulze (2026): study summary and source materials
