Paid clicks from ChatGPT's ads are worth a fraction of what advertisers are known to pay for them, according to a working paper by Maximilian Kaiser of the Vienna University of Economics and Business (WU Vienna) and Christian Schulze of the Frankfurt School of Finance & Management. The study, "Can Advertising Fund Free AI? Evidence from ChatGPT's International Advertising Rollout," is dated October 2026. It followed 1,898 online shops through OpenAI's country-by-country ad launch and found that in US traffic, a visit from a paid ChatGPT ad generated $0.89 in revenue, against $2.45 for a visit from a paid Google search ad.

The paper, version 1 and under review, gives no specific day of release. Its accompanying fact sheet opens with a blunt line: "ChatGPT's ads don't pay. Fixing them could cost ChatGPT its users." Measured at paid search's rate of return, the authors calculate that a ChatGPT click is worth about $0.26 to an advertiser. The largest publicly known account paid $1.72. Meanwhile, the arrival of ads coincided with an estimated 6 to 8% drop in the free, organic referrals ChatGPT sends to shops, although the visitors who still came bought at the same rate as before.

The study sets those findings against OpenAI's own economics. Advertising currently brings in about $1 per free user a year, according to the paper, while serving a free user costs about $5 in inference alone, and OpenAI's projections for 2030 imply about $40 per free user. "A Google competitor is not yet in sight," the abstract concludes.

In Short

ChatGPT started showing ads below its answers in February 2026, and two researchers used sales data from almost 1,900 online shops to see whether those ads work. They found that people who clicked a ChatGPT ad bought much less than people who clicked a Google search ad, so the ads are worth far less than advertisers are paying for them. Unless OpenAI finds a way to make each ad click more valuable, ads alone are unlikely to pay for free ChatGPT, and the obvious fix, putting ads inside the answers, could make people trust ChatGPT less.

How the study was built

The ad format under study is a labeled sponsored card placed below ChatGPT's conversational answer, shown only to users on the free tier and the low-priced Go tier. The design rests on the order in which OpenAI switched it on. According to the paper, ads launched in the United States on February 9, 2026; in Australia, Canada and New Zealand on April 16; in the United Kingdom on June 4; in Japan and Korea on June 19; and in Brazil and Mexico on August 4 and August 6 respectively. In Brazil and Mexico, the paper dates treatment from pilots that began showing ads to users; OpenAI opened both markets to advertisers on August 17, when ChatGPT ads reached nine markets. The European Union remained without ads for the whole observation window. OpenAI announced the EU expansion on August 15 - the paper cites PPC Land's report of that announcement for the date - and EU ads went live on August 24, two days after the data window closed on Saturday, August 22.

That sequence turns the 27 EU member states into a control group. Each rollout cohort is compared against the never-treated EU using the Sun-Abraham estimator, a method designed for staggered adoption that keeps countries already exposed to ads from serving as controls for countries treated later. Within each website, visitor country and device, the researchers compare organic ChatGPT referrals with organic Google search referrals week by week, so shocks that hit a shop or a market as a whole cancel out. Standard errors are clustered by website throughout.

The data covers 60 weeks, from June 29, 2025 to August 22, 2026, and nearly 10 billion visitor sessions drawn from the Google Analytics 4 e-commerce tracking of the panel websites. According to the paper, it was provided free of charge by Grips Intelligence, a market-intelligence company. Two panels serve two questions. The organic referral panel contains 1,898 internationally active websites with about $16 billion in annual sales; it tests whether ads damaged ChatGPT as a product. The paid referral panel contains 387 websites that received at least one click from a paid ChatGPT ad; 102 of them have complete financial tracking, and for 31 of those 102, linked Google Ads records supply advertising costs.

Over the full window, 22.4 million sessions in the organic panel came from organic ChatGPT referrals, against 0.45 million paid ChatGPT referrals in the paid panel. Organic ChatGPT accounted for 0.23% of all visits to the organic panel over the 60 weeks and 0.40% in the final four weeks. Paid ChatGPT sessions were far scarcer: 0.007% of all visits to the paid panel over the full window and 0.030% in its final four weeks. Organic ChatGPT converted at 1.99% over the full window; before the rollout, its conversion rate approximately matched that of organic search, according to the paper.

The tracking has limits that the paper states plainly. The data captures the last click that brought a visitor to a site. It does not reveal prompts, the set of recommendations ChatGPT showed, ad impressions or the subscription tier of the user. Treatment is therefore defined at market level: the estimates measure the effect of making ads available in a country, not the effect on an individual known to have seen one.

Channel tagging required one adjustment: in June 2026 Google Analytics reclassified ChatGPT referrals under an ai-assistant medium, and the researchers count both tag generations to keep the series continuous. Paid ChatGPT traffic is identified almost entirely by the chatgpt.com / cpc source and medium.

Organic referrals fell, buyers did not

The first question is whether advertising diminished ChatGPT's value to users. The researchers observe that value indirectly, through the AI referral traffic ChatGPT's free recommendations send to shops. Damage would show up either as fewer referred visitors or as visitors who buy less often.

On volume, organic ChatGPT referrals changed by -5.9% averaged over the first 13 weeks after launch, with a 95% interval from -12.6% to +1.2% and a p-value of .103. That interval includes zero. The paper adds a placebo test: in 100 simulated rollouts among EU countries where no ads launched, declines at least that large arose 10 times.

The estimate grows once faster ChatGPT adoption outside Europe is taken into account. According to OpenAI's own regional adoption index, cited in the paper, North America's weekly active users outgrew Europe's by 5.7% between January 2025 and January 2026, and by 11.6% at an annual rate from July 2025 to June 2026. Removing a 6% annual growth differential moves the estimate to -7.2% (interval -13.8% to -0.2%); removing 12% moves it to -8.4% (interval -14.9% to -1.5%). The interval excludes zero once the differential exceeds about 5.3% a year. The headline "6 to 8%" in the paper and fact sheet is, in the authors' own words in the appendix, "a range across assumptions, not an interval."

Conversion did not move. Organic ChatGPT conversion changed by -0.02 percentage points from a pre-launch baseline of 2.21%, with a p-value of .648 - a decline of about 1% whose interval comfortably includes zero. In 500 placebo rollouts, declines at least that large arose 187 times. The 95% interval rules out an immediate average decline larger than 6% of baseline; the placebo distribution rules out one larger than 10%.

Those figures describe all organic referrals from treated countries, including those from paying users who never see ads. Plus and Pro subscribers are about 5% of users but, according to figures from OpenAI's chief financial officer cited in the paper, use ChatGPT three (Plus) to eleven (Pro) times as much as free users. If paying users average five times the usage, ad-eligible users account for about 79% of referral sessions; at ten times, about 66%. Rescaled to those users, the bounds widen: declines of at most 7 to 12% in conversion and 16 to 19% in volume at the 79% share, and 9 to 15% and 19 to 23% at the 66% share.

Robustness checks leave the picture intact, though not without blemishes. Joint pre-trend tests give p = .458 for conversion and p = .109 for volume, but four of the five cohorts' own volume paths reject parallel trends, including the United States at p < .001, and the pooled test itself rejects in 49 of 100 placebo rollouts. The United Kingdom's conversion path alone gives p < .001; dropping the UK moves the pooled conversion estimate by a hundredth of a percentage point. Dropping the US cohort leaves volume at -5.8%. Dropping the largest websites, individually or the ten largest together, keeps volume between -5.1% and -8.8%.

The paper also notes how unevenly the cohorts contributed. The United States carries 61% of the post-launch weight in the conversion model, the United Kingdom 18%, Australia, Canada and New Zealand 16%, Japan and Korea 4%, and Brazil and Mexico 0.5%. Brazil and Mexico have only two post-launch weeks inside the window. By cohort, volume changes ranged from -23.4% (Brazil and Mexico, two weeks) to +13.6% (Japan and Korea, nine weeks); only the Brazil and Mexico figure is distinguishable from zero. "Because observed ad activity concentrated in the United States and United Kingdom, the pooled estimates describe the early rollout rather than five equally intense rollout cohorts," the paper states.

The researchers' conclusion on this first condition is measured. "Advertising has not substantially diminished the value of the product," according to the paper. "If conversion is read as a proxy for trust in ChatGPT's advice, the introduction of ads left that trust intact."

The second question is whether the ads are worth what they cost. The paper is explicit that this part is descriptive rather than causal: advertisers chose to buy ChatGPT ads, so they were not randomly assigned. The benchmark is paid Google search, which the paper describes as "the channel these same websites already buy for similar commercial intent, at prices the market has tested for decades."

In US traffic, where ad delivery concentrated (154,000 of 194,000 paid ChatGPT sessions in the financial sample), a paid ChatGPT visit generated $0.89 in revenue against $2.45 for a paid Google search visit, a 64% shortfall. Paid ChatGPT visitors converted at 0.37% against 1.17%. Within the same shops, adjusted for website, device and week, the conversion gap is -1.87 log-odds (p = .005), and paid ChatGPT orders are $74 smaller (p < .001).

Pooled across advertisers without adjustment, the paid ChatGPT basket is actually larger, $240 against $209, because advertisers differ in basket size; the paper quotes the within-website comparison.

The shortfall is broad. Of the 87 advertisers buying both paid channels in US traffic, 84 (97%) earn less per paid ChatGPT visit than per paid search visit and 85 (98%) convert worse. Seventy-three of the 87 - 84% - recorded no paid ChatGPT transaction at all, despite 18,726 paid ChatGPT sessions between them. The remaining 14 websites carry all 570 transactions. Had each advertiser's ChatGPT visits converted at its own paid search rate, the paper calculates, about 42 of the 87 would show no purchase by chance, not 73.

Concentration runs deep. One advertiser accounts for 66% of all paid ChatGPT purchases in the financial sample and 76% of US paid ChatGPT revenue. Removing it makes the channel look worse, not better: the conversion gap deepens to -2.83 log-odds and the revenue-per-session gap to -$4.31. Across 102 re-estimates, each dropping one advertiser, the largest p-value is .015 for conversion and .007 for revenue per session.

Measured against direct traffic, paid ChatGPT is the only channel that falls short on all six metrics. Among the 102 financially tracked advertisers, its visitors convert far less often (log-odds -1.82, p = .011), place orders $79 smaller and generate $2.86 less revenue per session. Google's paid search does not differ detectably from direct traffic on any of the three financial measures. On the wider panel of 387 websites, paid ChatGPT visitors bounce more often, view 1.4 fewer pages and stay 104 seconds less than direct visitors. Paid search, organic search and organic ChatGPT visitors all bounce less, view more pages and stay longer than direct visitors.

The contrast between ChatGPT's two channels is stark. Organic ChatGPT visitors in the financial sample convert at 1.31% across all countries and generate $2.48 per session; paid ChatGPT visitors convert at 0.31% and generate $0.73. In US traffic, organic ChatGPT generated $3.27 per session, the highest of the five channels.

The $0.26 parity price

The paper's most quotable number comes from a simple calculation. A paid search click returns $2.45 on average. For the 16 accounts in the cost sample whose paid search traffic comes at least 80% from the United States, the median search cost per click over the window was $0.72. That is close to the $1.01 median account price on Google's search network across 40,183 accounts in the months before the ad launch, according to the paper.

A paid ChatGPT visit returns $0.89. At the same rate of return as search, a ChatGPT click would therefore be worth $0.72 multiplied by 0.89/2.45: $0.262, rounded to $0.26. The researchers call this the parity price - the price at which a ChatGPT click returns as much per dollar as a search click. Resampling advertisers 2,000 times gives a range of $0.05 to $0.57; one advertiser carries 76% of US paid ChatGPT revenue, so resamples with and without it differ widely.

Actual prices sit well above that range. According to the paper, the largest publicly known account, reported by Opascope in June 2026, paid $1.72 per ChatGPT click - 6.57 times the parity price, and 2.38 times the median search click. The authors describe that comparison as conservative: $1.72 sits toward the low end of what practitioners report and below the starting maximum bid of $3 to $5 per click that OpenAI recommends - a figure OpenAI gave when Ads Manager opened to all US businesses with cost-per-click bidding on May 5. Taken together, the paper states, "ChatGPT ads offer more than six times worse value for money than Google Search ads."

Two caveats apply, and the paper names both. The parity price assumes similar profit margins in both channels, and it counts only the tracked visit, not purchases that happen later through another route. To probe the second, the researchers matched 305 advertisers to similar websites that never bought ChatGPT ads and compared traffic in the 12 weeks before and after each advertiser's first paid click. Organic ChatGPT traffic rose an estimated 8%, not distinguishable from zero; direct, organic search and paid search traffic showed no detectable change. With a smallest detectable increase of about 50%, the design "can rule out only large lifts, not moderate ones," the appendix acknowledges.

The public success story does not travel well. Opascope's account reported about $60,000 of spend over 15 days, a 2.35% conversion rate and $1.49 of revenue per dollar spent, about $2.56 per visit, according to the paper. In the panel of 102 advertisers, the median website records 0.00% conversion and $0.00 per paid ChatGPT visit; the 90th percentile reaches 0.14% and $0.25. One website matches the Opascope account on both measures, and it does so on 9 paid sessions. The panel's own largest paid advertiser, with 52,232 sessions, converts at 0.76% and earns $1.99 per visit. The best-known public success story, the paper states, "is the exception, not the rule."

The financing gap

The study's title asks whether advertising can fund free AI. Its answer rests on arithmetic that the authors label illustrative, built from public figures and one estimate of their own.

Start with revenue. OpenAI reported $1 billion in projected annualized advertising revenue on August 31, less than 200 days after launch. With more than 1 billion weekly users and about 95% of them not paying, the free base is about 950 million, giving $1.05 per free user per year. That average includes users in markets where no ads ran. The cost side is roughly $5 per free user per year in inference cost, a figure the paper attributes to OpenAI's own numbers as reported by The Information; the appendix notes it is an estimate of inference alone, "not the full cost of serving a free user."

Then the target. OpenAI's projections, as reported by The Information in April 2026, anticipate $102 billion in advertising revenue by 2030. Spread over the roughly 2.52 billion free users the company's forecasts imply for that year, that is $40.48 per free user, rounded to $40.

To translate those targets into ads, the paper multiplies four inputs: commercial prompts, ads per commercial prompt, click-through rate and value per click. A free user sends about seven prompts a day, according to OpenAI's chief financial officer, and a fifth of prompts carry commercial intent, according to OpenAI's comments at Cannes reported by Marketing Dive. That gives 511 commercial prompts a year. At a 0.50% click-through rate - Similarweb's June 2026 estimate - and the $0.262 parity price, one ad on every commercial prompt earns $0.67 per free user per year.

Covering the $5 cost would take 7.5 ads per commercial prompt. Reaching $40 would take 59.8, which the paper rounds to about 60. Even at the most favorable click-through rate the paper found reported, 1.30% from SE Ranking's own campaigns, the $5 would take 2.9 ads per prompt and the $40 would take 23.0. At the lowest, 0.35% from one advertiser's account, the figures are 10.7 and 85.5. Click-through rates have risen since the paper's input was measured. Similarweb and Graphite reported in September that ChatGPT's worldwide weekly ad click rate had risen 112% since late March, with an all-time US rate of 0.73% between March 30 and September 16. At that rate and the same parity price, covering $5 would still take about 5.1 ads per commercial prompt and $40 about 41, by PPC Land's calculation using the paper's formula.

The current ad load, by the paper's calculation, is about 0.23 ad units per commercial prompt: $1 of revenue at $1.72 per click is 0.58 clicks per free user a year. That squares roughly with an SE Ranking sample that found ads on 25.9% of 50,006 commercial US prompts in July 2026. Other trackers measure exposure differently: Similarweb put the share of ChatGPT responses carrying a sponsored ad at 26% for June 2026 US desktop traffic, while Adthena found ads on 4.47% of all US queries it sampled between March and May, with retail taking 39% of ad items. "That is not a realistic scenario, so the value per click must improve," the authors write of the 60-ad case.

The fact sheet compresses the point into one sentence: "More advertisers and more clicks will not close the gap: each click has to become worth far more."

Moving ads into the answer

That leads to the dilemma the paper ends on. One reading of the results, according to the authors, is that ChatGPT clicks are worth little because the sponsored cards sit apart from the answer, where the product's value lies - and that the same separation is why organic referrals were largely spared.

Integrating ads into the answer could raise their value. The paper cites a preregistered experiment by Salvi, Cuevas and Horta Ribeiro, posted in April 2026, in which a model arguing for a sponsored product won a majority of choices, largely unnoticed even when labeled, while a neutrally described sponsored item moved choices no more than a search placement. Yet a 2025 study by Tang and colleagues found that users who become aware of ads embedded in an answer feel manipulated and voice concerns about bias. "Integrating ads into the answer would then buy value per click with users' trust, potentially improving Condition B while risking Condition A," the paper states.

The fact sheet frames the same trade-off for a general audience. "OpenAI needs more valuable ads. It could make them more valuable through better targeting or by moving them into the answer, but doing so would endanger the product even more."

The authors also draw a line to antitrust. The remedies Judge Amit Mehta imposed on Google in September 2025 rest partly, the paper argues, on the expectation that AI assistants will become advertising-financed competitors for Google's commercial queries. "For the industry, the advertising-financed competitor to Google that the search-case remedies count on is not yet in sight," according to the paper, "and more advertisers and clicks alone will not bring it, because unfavorable unit economics do not scale away."

What the study does not establish

Several limits shape how far the findings reach. The paper is a working paper, version 1, under review; it has not completed peer review. The window closed on August 22, 2026, before ads reached the EU and before most of the product changes OpenAI has made since. The paid evidence covers the first 28 weeks of advertising and rests on 0.45 million paid sessions, the large majority from the United States.

The two panels are different populations. The organic panel is predominantly retail: home and garden, fashion, electronics, health and sports together hold 55% of its websites. In the paid panel, services, software, education and finance make up about half the websites, and the five retail categories that lead the organic panel less than a third. Paid search results describe what those advertisers saw, not what every merchant would see.

Campaigns that never produced a click are invisible, because a website only counts as an advertiser from its first paid-tagged session. First-click dates can lag true campaign starts. Profit margins are assumed equal across channels. The engagement and cross-channel analyses can show patterns consistent with additional value but cannot separate it from the advertiser's own decision to advertise.

There is also the matter of the data source. According to the paper's disclosure, Maximilian Kaiser was previously employed by Grips Intelligence and now serves as its scientific adviser; beyond his salary and virtual company shares, the authors received no payment, equity or other consideration from the company. The research agreement gives Grips Intelligence "no right to review, approve, or delay publication," and the company, according to the paper, has no commercial relationship with OpenAI and does not sell ChatGPT advertising or measurement services tied to it. Christian Schulze declares no competing interests.

Why the findings matter for marketers

ChatGPT's ad business has been one of the most closely followed stories in search advertising this year, and much of the performance data published about it has come from OpenAI, its ad tech partners or measurement vendors. When OpenAI announced its ad test on January 16, it described clearly labeled sponsored units at the bottom of answers for Free and Go users in the US. By late March, Reuters reported more than $100 million in annualized revenue and more than 600 advertisers. On August 31, OpenAI said the business had passed a $1 billion annualized run rate, and on October 21 it said advertisers in 89 countries could buy ChatGPT ads and more than 100,000 had run campaigns since launch.

None of those milestones describe what a click returns to the business paying for it. That is the gap this paper tries to fill, from outside the platform and with merchants' own analytics. Its central claim is not that ChatGPT ads fail to sell; 14 of the 87 dual-channel advertisers recorded paid ChatGPT purchases, and the panel's largest advertiser earned $1.99 per visit. The claim is that, on average and in the first 28 weeks, the value per click sat far below the price, and that this gap is a problem of unit economics rather than scale.

The revenue projections that frame the paper have been debated before. In June, PPC Land examined Barclays' forecast of $102 billion in ChatGPT ad revenue by 2030, set against roughly $224 billion in Google Search and other revenue in 2025. Days later, EMARKETER put US ad revenue for ChatGPT and other standalone chatbots at just over $5 billion by 2030, with most AI ad spending appearing next to AI content rather than inside chatbots. The Kaiser and Schulze paper adds a bottom-up view: its $40-per-free-user target would require about 60 ads per commercial prompt at today's measured value per click.

The question of whether ads can carry an AI product has already been tested in bankruptcy court. In October, Character.ai's chief executive declared that the chatbot app had grown advertising revenue to about $2.5 million a month while losing about $8 million to $9 million a month, with compute taking a large share of costs. Character.ai and ChatGPT differ in scale, format and audience, and the paper does not discuss Character.ai. But the arithmetic is similar in shape: advertising revenue rose quickly, and the cost of generating answers stayed well ahead of it.

Vendor data tells a partly different story. Criteo has said that LLM-referred shoppers convert at rates approaching twice those of traditional search in several retail categories. The new paper is closer to that for organic referrals, which generated $3.27 per session in US traffic, the most of any channel measured. The divergence is in the paid channel.

For advertisers, the measurement limits matter: the data captures only the last click, and the matched analysis could rule out only large lifts in other channels. Conversion-optimized campaigns, which OpenAI began rolling out on June 5, fall inside the last 11 weeks of the window; the paper does not separate their performance. And the study window closed before ads reached Europe. When OpenAI told European users on August 15 that ads would arrive later that month, it said selection would not be personalised at launch, and advertisers would receive only aggregate performance data. Whether those conditions change the value of a click is outside the paper's scope.

For retailers, the organic findings may matter more. A 6 to 8% reduction in organic ChatGPT referrals, if it holds, is small in absolute traffic, but it lands on a channel whose share of visits in the panel reached 0.40% in the window's final four weeks, and which converts at rates close to organic search.

What the paper describes, in the end, is a choice OpenAI has not yet had to make. Its ads sit below the answer, and the authors suggest that this separation both protects the product and limits the ads' value. The fact sheet describes the result plainly: "Even these separated ads have moderately hurt ChatGPT as a product." The authors' closing sentence names the open question: "The central challenge for AI assistants is whether advertising can become valuable enough to finance the product without making the product less useful."

Timeline

Summary

Who: Maximilian Kaiser of the Vienna University of Economics and Business, who is also scientific adviser to data provider Grips Intelligence, and Christian Schulze of the Frankfurt School of Finance & Management, studying OpenAI's ChatGPT ads and the 1,898 online shops and 387 advertiser websites in their data.

What: A working paper, "Can Advertising Fund Free AI? Evidence from ChatGPT's International Advertising Rollout," finding that ads coincided with an estimated 6 to 8% fall in organic ChatGPT referrals with no change in conversion, that a paid ChatGPT visit in US traffic generated $0.89 against $2.45 for paid Google search, and that a ChatGPT click is worth about $0.26 at search's rate of return against a known price of $1.72.

When: The paper is dated October 2026, version 1, and is under review. Its data covers June 29, 2025 to August 22, 2026, spanning ad launches from February 9, 2026 in the United States to August 4 and 6 in Brazil and Mexico.

Where: Nine countries where ads launched - the United States, Australia, Canada, New Zealand, the United Kingdom, Japan, Korea, Brazil and Mexico - compared with the 27 EU member states, where ads arrived only after the window closed.

Why: OpenAI's advertising brings in about $1 per free user a year against roughly $5 in inference cost and a 2030 projection of about $40, and the authors argue that only more valuable ads can close that gap, while moving ads into ChatGPT's answers to make them more valuable risks the trust that keeps users coming back.