Strong AI referrals do not prove ChatGPT ads work
An unpaid ChatGPT recommendation can bring a retailer a valuable customer even when a paid placement delivers a weak return. So an advertiser considering ChatGPT Ads needs to distinguish the two routes before setting a budget. OpenAI earns revenue from selling those placements, so it has a commercial interest in increasing advertiser spending while preserving the recommendation experience that attracts users.
ChatGPT makes that distinction visible: users on its Free or Go tier receive buying advice with a separately labeled “Sponsored” card below it. Because the card sits outside the recommendation, buying that space gives an advertiser a different route to the customer from earning an unpaid recommendation. Performance on one route cannot establish what the other is worth.
Paid clicks generate weaker measured returns
The paid-traffic comparison comes from researchers associated with the Frankfurt School of Finance & Management and the Vienna University of Economics and Business. Their working paper, still under review, analyzed 1,898 ecommerce sites with $16 billion in combined annual sales and nearly 10 billion visits over 60 weeks, from June 2025 through August 2026. Those totals cover the ecommerce dataset across acquisition channels.
Within that dataset, paid ChatGPT visits generated average revenue per session of 89 cents, compared with $2.45 for paid Google Search. Revenue per session measures the sales revenue associated with each visit; conversion rate measures the share of visits that led to a purchase. Conversion rates were 0.37% for paid ChatGPT and 1.17% for paid Google Search. Together, these measures show fewer purchases and less sales revenue from an average paid ChatGPT visit.
The weakness extended beyond paid search. The researchers also found that paid ChatGPT visitors underperformed visitors arriving through Google ads, Google search, direct traffic and unpaid ChatGPT recommendations. These are different ways of acquiring customers, so the broader finding puts paid ChatGPT’s results in context without assigning the paid Google Search figures to every other channel.
At the advertiser level, the average conversion rate concealed how often campaigns recorded no purchases at all. Among 102 advertisers in the study, 82% recorded no purchases from their ChatGPT ads. For those accounts, reporting showed advertising visits without sales revenue to offset their acquisition cost. The percentage applies to that advertiser sample.[
Whether those visits offer a competitive return also depends on their price. The researchers first compared conversion and revenue per session, then used publicly reported advertising prices to estimate sales revenue per advertising dollar. On that basis, paid Google Search generated more than six times the revenue per advertising dollar that ChatGPT produced. This connects visitor behavior to the amount an advertiser spends to obtain it.
The researchers estimated that ChatGPT would need to cost about 26 cents per click to match paid search’s return. The largest publicly known account in their data paid $1.72 per click, giving an account-level comparison with the calculated price. One figure estimates the price needed to match returns under the researchers’ comparison; the other records what an account paid. A visit producing relatively little revenue can still be attractive at a low enough price, which makes the gap commercially important.
A project in mind?
Schedule a 30-minute meeting with us.
Senior experts helping you move faster across product, engineering, cloud & AI.
Attribution affects how advertisers read the results
The spending calculation also depends on whether purchases receive the correct advertising credit. Farhad Divecha, group CEO at Accuracast, an advertising agency with a commercial interest in helping clients measure and improve campaigns, told Digiday: “We continue to see discrepancies there,” and explained: “Some are teething problems because it’s a new system that has to be set up on advertiser systems, but some problems are down to attribution being messed up.” Attribution means assigning a purchase to the advertising interaction that receives credit for it.
Divecha identifies problems with both advertiser-side setup and the assignment of purchase credit. A purchase missing from campaign reporting may still have happened, so the no-purchase finding describes recorded outcomes. Fixing those problems could change measured performance, but the effect depends on how many purchases were misassigned and in which direction. Advertisers therefore need to assess recorded returns alongside the reliability of the system assigning them.
Unpaid AI referrals measure a different opportunity
The same working paper provides a separate reason to distinguish paid and unpaid traffic. Before advertising began, unpaid ChatGPT referrals converted at 2.21%, compared with the paid-visit rate of 0.37%. Because the unpaid figure comes from an earlier period, the rates are not a controlled comparison from the same time. Even so, they show the risk of using past unpaid-referral performance to set expectations for a paid placement.
Broader retail measurements also show valuable AI referrals. Adobe Analytics, whose analytics business benefits from demand for measuring digital customer behavior, found that AI referrals to U.S. retail sites generated 83% more revenue per visit than non-AI traffic in September. Those visitors also converted 90% better, spent 73% more time on retail sites and viewed 25% more pages. The results combine stronger purchase outcomes with greater engagement.
But Adobe’s definition covers AI referrals broadly and does not isolate paid ChatGPT clicks. More time and more pages describe browsing behavior, while conversion and revenue per visit connect visits to sales. These findings help explain interest in AI as a route to customers, but they cannot determine what an advertiser should pay for a ChatGPT ad click. Ecommerce teams need to keep the acquisition route and measurement period attached to each result.
Advertiser growth measures participation
Interest in buying those clicks is growing, even as the paid-performance study raises questions about returns. Adthena, an advertising-intelligence business with a commercial interest in demand for campaign monitoring, tracked 1,082 brands advertising across the U.S., U.K., Australia and EU in June. Its tracked count rose to 3,741 in July and 6,333 in August. Those figures show rapid growth in observed participation.
New Street Research’s estimates also point to increasing adoption, using a different measure. It estimated about 20,000 ChatGPT advertisers at the end of August and roughly 60,000 by early October. Because those estimates differ from Adthena’s tracked brand counts, they should be read as separate series. Both describe participation, while campaign returns require sales and spending measurements.
Participation is also spreading across spending categories. Sensor Tower, a market-intelligence provider with a commercial interest in demand for advertising data, supplied figures cited by Digiday showing shopping’s share of U.S. ChatGPT ad spending falling from 38% in April to 20% in September. Over the same period, software rose from 13% to 18%, financial services from 2% to 13%, and travel and tourism from 5% to 9%. Because these are spending shares, they show a broader category mix without establishing changes in each category’s absolute spending or return.
OpenAI’s revenue ambitions depend on click value
A larger advertiser base matters to OpenAI, but the researchers’ estimates show why participation alone cannot settle the economics. They estimate that current advertising generates about $1 per free user annually, against roughly $5 to serve a free user. On those estimates, advertising revenue falls below the cost of serving that user. Closing the gap requires a substantial increase in revenue per user.
OpenAI’s projections put advertising revenue at $102 billion by 2030, equivalent to about $40 per free user. The researchers conclude that adding advertisers and clicks will be insufficient to close the funding gap unless each click becomes materially more valuable. Advertisers’ ability to pay depends on what the traffic earns them, which ties OpenAI’s projected revenue to campaign performance. More available visits alone cannot establish that advertisers can profitably buy them at prices that support those projections.
OpenAI reports that ChatGPT Ads reached a $1 billion annualized revenue run rate less than 200 days after launch. An annualized run rate expresses the yearly revenue pace implied by activity at a particular point, so it differs from cumulative advertising revenue already collected. The milestone describes growth in OpenAI’s advertising business, while advertiser returns remain a separate measurement.
Query volume measures scale
Audience size offers another measure of the business’s potential, but query definitions affect the comparison. OpenAI claims ChatGPT handles roughly 20% as many “search-like” queries as Google. Applying that share to Google’s last disclosed annual search volume implies more than 1 trillion such ChatGPT queries annually. The total is an extrapolation from the claimed share, and differences in what counts as a search limit comparison with Google Search.
SEOClarity, a search-optimization software company with a commercial interest in how businesses measure search activity, estimates a smaller volume. It puts ChatGPT at about 7.5% of Google’s search volume, implying approximately 375 billion queries annually. The estimates differ, and each depends on its measurement approach. Even a precise query count would leave advertisers needing to establish how much revenue purchased visits generate at the price charged.
The referral decline raises a product-design question
Beyond audience size, the researchers examined whether introducing ads coincided with changes in the number of ecommerce visitors ChatGPT sent. In their study design, ads began in the U.S. and subsequently expanded to eight additional countries, while ad-free EU markets served as controls: comparison markets used to estimate changes without advertising. They compared post-launch referral changes in markets with ads against changes in those control markets. The analysis concerns referral volume, extending the question beyond returns on purchased clicks.
The researchers’ broad finding was a 6% to 8% decline in visitors sent to the ecommerce sites after advertising appeared. Their initial estimate was a 5.9% decline relative to markets without ads during the first 13 weeks. But it was not statistically significant, meaning the statistical test did not distinguish it clearly enough from chance variation. The initial result needs to be read with that qualification.
The researchers then adjusted for ChatGPT usage growing faster in North America than in Europe. Under that adjustment’s assumptions, the estimated decline ranged from 7.2% to 8.4% and became statistically significant. So the stronger statistical result depends on how the analysis accounts for unequal regional usage growth. Both the estimated decline and its statistical significance change with that adjustment.
The observed decline concerned referral numbers, while unpaid visitors who still arrived converted at approximately their previous rate. Fewer visits and weaker conversion would imply different product problems, so the distinction matters when assessing what changed. The analysis identifies lower referral volume without establishing its cause. It also leaves open the possibility that paid performance could improve while unpaid recommendations retain their value.
To improve paid performance, the researchers suggest better targeting and placing ads more directly within answers as ways to increase the value of advertising space. Better targeting could make sponsored offers more relevant, while placement within answers would change their relationship to the buying advice itself. The current separation between recommendations and the sponsored card gives context to that product-design proposal. Moving ads into answers could put pressure on the experience supporting unpaid recommendations, but that remains a proposed tradeoff rather than an observed result of an experiment with those placements.
The bottom line
For business leaders, the decision is not whether AI can bring valuable customers. It is whether paid ChatGPT placements can acquire them at a price that supports the business. Strong unpaid referrals, rising advertiser participation and OpenAI’s revenue growth do not answer that question.
The current evidence supports measured testing rather than broad budget shifts. Advertisers should evaluate paid ChatGPT traffic separately from unpaid recommendations, verify purchase attribution and compare acquisition costs with established channels. Revenue per visit is useful, but a spending decision also needs to account for margins and the cost of each click. Attribution problems warrant investigation, not an assumption that unrecorded sales will make a campaign profitable.
For OpenAI, improving advertiser returns may require changes to targeting, pricing or placement. Those choices also need to preserve the recommendation experience that makes ChatGPT valuable to users. The commercial test is whether paid placements can deliver competitive returns without weakening that experience. Until the evidence is stronger, budget growth should follow verified results rather than platform momentum.
A project in mind?
Schedule a 30-minute meeting with us.
Senior experts helping you move faster across product, engineering, cloud & AI.


