Published: November 6, 2026
Last Updated: November 6, 2026
ChatGPT ads generated less revenue per visit and converted less often than Google Search ads in an October 2026 working paper. In US traffic, the researchers reported $0.89 per paid ChatGPT session against $2.45 for paid Google Search, with conversion rates of 0.37% and 1.17%. For a local service business, our advice is to protect advertising that produces profitable work, keep building organic visibility, and treat paid ChatGPT as a measured experiment rather than a replacement. Read the paper.
A roofing company does not need to win the race into every new ad platform. It needs the right calls, estimates that become jobs, and enough margin left after the crew gets paid. A full inbox of sales pitches about the next big thing does not change that math.
The paper deserves your attention. It also deserves a careful reading. It does not prove that every ChatGPT campaign fails, that Google always wins, or that AI visibility is a waste of money. Buying an ad beside an answer and being cited inside an answer are different jobs.
Paid ChatGPT traffic delivered less tracked revenue per visit and a lower purchase conversion rate than paid Google Search traffic in the US comparison.
Maximilian Kaiser of the Vienna University of Economics and Business and Christian Schulze of the Frankfurt School of Finance & Management wrote Can Advertising Fund Free AI? Evidence from ChatGPT’s International Advertising Rollout. It is a working paper under review and has not completed peer review. See the research overview and Search Engine Journal’s review of its methods.
The overall research covers 60 weeks from June 29, 2025 through August 22, 2026, with nearly 10 billion sessions across an organic referral panel of 1,898 websites. But the paid revenue comparison uses a much smaller financial sample of 102 websites. Those are not interchangeable sample sizes. See the sample definitions.
| Measure | ChatGPT ads | Google Search ads | What the number describes |
|---|---|---|---|
| Revenue per session | $0.89 | $2.45 | Pooled US traffic in the financial sample |
| Conversion rate | 0.37% | 1.17% | Tracked transactions per session in US traffic |
| Advertisers earning less per ChatGPT visit | 84 of 87 | Comparison channel | Websites buying both channels in US traffic |
| Advertisers with no ChatGPT purchase | 73 of 87 | Not the measure reported | The same US comparison group |
The authors describe the revenue gap as a 64% shortfall. Their adjusted comparisons also found lower conversion and revenue for paid ChatGPT after accounting for website, device and week differences. This was not simply a comparison of unrelated headline averages. See Table J.1 and Table K.1.
Search Engine Land led with the revenue shortfall. Digiday put the findings beside the growth in advertisers. Both are useful context. For the owner of an HVAC company, the next question is simpler. Does the channel bring you profitable work?
Only if the price is low enough relative to the value that visit produces. A cheap click can still be an expensive customer.
The researchers estimated that a ChatGPT click would need to cost about $0.26 to match Google Search on tracked revenue per advertising dollar. They used a $0.72 median Google Search click price from 16 accounts dominated by US traffic, scaled by the difference in revenue per session. The estimated parity price ranged from $0.05 to $0.57 when advertisers were resampled. See the pricing calculation.
The $1.72 ChatGPT click price discussed in the paper comes from a publicly disclosed outside account. It is not the average price paid by the study’s advertisers. Comparing that price with the estimated $0.26 explains the authors’ conclusion about more than six times worse value for money. It does not give every advertiser a measured return on ad spend.
Revenue per visit is not revenue per dollar spent. Neither is profit. The parity calculation assumes similar margins and counts the tracked visit. Do not turn a comparison built from these inputs into a promise about your own account.
There is no measured ChatGPT click through rate for the study panel because the researchers could not observe ChatGPT ad impressions. The 0.50% rate used elsewhere in the paper is an outside estimate for its platform financing arithmetic, not a measured result from these advertisers. See the financing assumptions.
No. The paid channel comparison describes observed results from advertisers who chose to participate. It is not a randomized test of local contractor campaigns.
The data came from website GA4 tracking supplied by Grips Intelligence. Kaiser is a scientific advisor to the company and previously worked there. The paper discloses that relationship and says Grips had no role in the research decisions or writing. Read the disclosure.
The authors could not see prompts, ad impressions or the full customer journey. A later purchase through another route may not appear as revenue from the original visit. Campaigns that never generated a click are invisible. The results also describe an early rollout, not every later version of the advertising product. Read what the data can show.
Most important for our clients, this is not a representative benchmark for qualified roofing calls, HVAC installations or signed remodeling contracts. A store purchase and a request for an estimate are not the same conversion. The paper gives you a reason to demand evidence, not permission to paste its rates into your own forecast.
No. Paid ChatGPT placements and unpaid recommendations are different channels. Weak paid ad results do not establish that earning an organic citation has no value.
In the same financial sample’s US traffic, organic ChatGPT referrals generated $3.27 per session, compared with $0.89 for paid ChatGPT visits. That is a descriptive comparison, not proof that an answer engine optimization program will generate a particular return. But it is a good reason not to put every AI visit in the same bucket. See the separate channel results.
The organic story also has limits. The estimated referral volume change after ads arrived was a 5.9% decline, with an interval that included no decline. The larger estimates often summarized as 6 to 8% depend on assumptions about regional usage growth. The researchers found no detectable change in conversion among the organic visitors who still arrived. See the organic results and uncertainty.
For Prager, the practical distinction is buying space versus building something worth citing. An ad buys a placement. Our AEO work focuses on making the business’s expertise clear and accessible enough to be useful in an answer. Neither route guarantees that a homeowner will call.
Google’s own guidance for AI Overviews and AI Mode says existing SEO fundamentals remain relevant. A supporting page must be indexed and eligible for a search snippet. There is no special schema requirement and no guarantee of inclusion.
That makes organic SEO and AEO connected work, not rival subscriptions. A useful page about whether a roof needs repair or replacement can help a homeowner, support search visibility and give an answer engine a clear explanation to reference. The page still needs real expertise. A stack of generic AI paragraphs is not a strategy.
Protect the channels producing profitable jobs, fund useful organic content and technical improvements, and keep unproven ad experiments separate. There is no evidence here for a universal percentage split.
Our approach is to give each dollar a job before giving it a platform. A contractor with open crews and a working campaign has a different problem from a remodeler with a full schedule and a weak website.
Do not fund an experiment by quietly draining the campaign that pays the bills. Do not protect an old campaign forever either. If the new channel earns profitable customers under a fair comparison, it can earn more budget. The platform name does not get a vote.
Track qualified inquiries through estimates, booked work and collected revenue. A form submission is a starting point, not the finished result.
A contractor can look successful in an ad dashboard while the office spends the day rejecting calls outside the service area. Count the wrong event and the report gets prettier while the business gets busier with the wrong work.
A roof replacement and an emergency furnace repair do not have the same decision cycle. Treating them as interchangeable can make the fastest closing service look like the best advertising platform.
Check your analytics before interpreting a rise in Direct traffic as proof of AI performance. An unknown referral is not a verified AI visit. Keep confirmed AI referrals, paid traffic and unattributed visits separate, and record source information with the lead wherever possible.
Marketing Brew’s reporting on measurement updates shows that the tools continue to change. That is a reason to keep checking your setup, not a reason to assume missing revenue must exist.
Demand proof from paid AI advertising without abandoning the work that helps your business earn organic visibility.
That is how we approach paid search, SEO and AEO at Prager. Not as a contest between platforms, but as connected work that has to make sense for your business. If the reporting stops at clicks, the conversation stopped too soon.
The main distinctions are the sample, the conversion being measured and whether you are discussing paid placements or organic recommendations.
In the working paper’s US comparison, paid ChatGPT visits generated $0.89 per session versus $2.45 for paid Google Search, and converted at 0.37% versus 1.17%. These observed results do not predict every advertiser’s performance.
No. The organic referral panel contained 1,898 websites. The paid referral panel contained 387 websites, and the financial sample used for revenue and conversion analysis contained 102 websites.
No. The researchers did not observe ChatGPT ad impressions. The 0.50% click through rate in the paper’s financing calculation is an outside estimate, not a measured rate from its advertiser panel.
Not automatically. Keep the experiment separate from proven campaigns, define a qualified lead before launch and judge results through booked work. The paper is not a representative benchmark for contractor campaigns.
No. An ad is a paid placement. A citation is a reference in an answer. The study reports these as separate traffic channels and does not establish the return from an answer engine optimization program.
No. It does not test an ideal budget allocation. Prager recommends protecting profitable campaigns, improving the website that supports organic discovery and reserving a separate amount for experiments your business can afford.
FAQ findings are drawn from the paper’s sample definitions, channel results and financing assumptions. Budget recommendations are Prager’s interpretation, not findings from a budget allocation experiment.
