Where AI ads are today

1.2 Return

Incremental return is the missing measurement

Every number about assistant ad performance currently comes from one of three places: the platform itself, a vendor that sells services in the category, or a single working paper. None comes from a controlled test of whether the ads produce sales that would not otherwise have happened.

A vendor panel measured click-through by market

A panel run by Similarweb and the growth agency Graphite tracked ChatGPT ad exchanges from March 30 to September 16, 2026. Click-through ran highest in Latin America and lowest in English-speaking markets, though the Latin American markets launched months later, so the comparison is not like-for-like. Worldwide click-through rose 112% across the period.⁴⁰

Click-through per ad-bearing exchange, pooled across advertisers. Bar length scaled to 2.0%.

One working paper measured what the clicks were worth

Researchers at WU Vienna and the Frankfurt School compared US advertisers buying both paid ChatGPT and paid search, using analytics panel data from Grips Intelligence. A paid ChatGPT visit produced about a third of the revenue of a paid search visit and converted at about a third of the rate.⁶

Results were concentrated. Of 87 advertisers buying both channels, 73 recorded no paid ChatGPT transaction during the window, and a single advertiser accounted for 66% of paid ChatGPT purchases. Organic ChatGPT referrals to the sites in the panel fell 5.9% after ads launched, an estimate whose confidence interval includes zero. The paper is not peer reviewed.

2.2 Creative

The platform now writes part of the ad

OpenAI switched on “creative enhancements” for new and existing self-serve campaigns on October 30, 2026. The system can rewrite headlines and descriptions, translate ad text, resize images and change destination URLs. Variations can be generated in real time from the conversation, and enhanced versions “can run automatically without requiring you to review or approve each version.” Advertisers can switch the feature off for an account or a campaign.⁹

Google is testing Conversational Discovery ads in AI Mode, with creative Gemini writes to answer the person’s question. In AI Overviews, some sponsored placements already carry short Gemini-written descriptions that advertisers cannot edit or opt out of. AI Mode itself cannot be excluded from the campaigns that feed it.^2, 7

In financial services, health, education and other regulated categories, this creates compliance exposure. The approved claim and the served claim can now differ without anyone at the advertiser seeing the served version.

3.1 Designs

The research assumes the ad is written into the answer

Shipped formats keep ads outside the generated text. Most academic and industry-lab work on selling ads in assistants assumes the advertiser’s message is generated into the answer, with the auction pricing what the model writes.

A second design, presented at NeurIPS 2024, runs a separate auction for each segment of an answer, such as a sentence or a paragraph, and prices each placement per click. The simulations put numbers on a tradeoff for any platform that sells ads inside answers. Auctioning one ad per segment earned 0.371 in expected revenue against 0.238 for bundling several ads into one placement, about 36% less. The bundled version produced answers closer to what the model would have written with no ads at all.¹¹

Every performance figure in this literature is simulated. None of these designs has been tested on a live assistant.

3.2 Users

Assistant persuasion nearly tripled sponsored choices in a controlled test

The most direct test of what happens when an assistant has a commercial interest comes from a preregistered Princeton experiment. In it, 2,012 US participants chose a free book from about 5,500 titles, either through a search interface or through a chat assistant running one of five leading models. One title in five was secretly marked as sponsored, so a random pick lands on a sponsored title 20% of the time.¹²

Placement alone, in search or in chat, moved choices only a few points above chance. Persuasive language from the assistant nearly tripled the search figure. The assistants used ordinary sales technique, talking up the sponsored option and hedging on the alternatives.

3.3 Organic answers

Organic answers can already be steered without buying an ad

• Adding quotations, statistics and citations to a source raised its visibility in generated answers by up to about 40%, with the largest gains for lower-ranked sources.⁵⁵
• In 2024, text written to manipulate preferences made Bing Copilot recommend a test product about 2.5 times as often as a comparable one, and lifted a fictitious camera from 34.0% to 59.4% of recommendations. When several sellers tried the same tactic, all of them lost visibility. Microsoft said it was adding protections.⁵⁶
• Across repeated runs of the same product query, ChatGPT’s recommended product sets overlapped by 0.178 on average, Gemini’s by 0.287 and Google AI Overviews’ by 0.421. Among ChatGPT responses that named products, 73% labeled one “best.”⁵⁷
• Models from OpenAI, Anthropic and Google ranked their own companies’ products about 0.2 positions higher than benchmark performance justified.⁵⁸

Share of answer, the rate at which a brand appears in responses, is volatile before any ad runs. A test that compares organic mentions before and after an ad campaign needs a large, fixed set of prompts and a control period long enough to absorb that noise.

4.1 Precedent

Paid search faced the same questions twenty years earlier

Between 1998 and 2015, search advertising worked through whether people can tell paid results from organic ones, how prominent the label must be, and whether platform-reported returns hold up under testing.

The FTC set the disclosure standard in 2002

In 2002 the FTC told search engines to “clearly and conspicuously disclose” paid placement, using terms that convey the ranking is paid for.⁵⁹ In 2013 it warned that paid results had become “less distinguishable as advertising,” and stated that its guidance applies “regardless of the precise form that search takes now or in the future.”⁶⁰ FTC business guidance published in 2023 extended the principle to generative AI: output should distinguish clearly between what is organic and what is paid.⁶¹

Google’s ad label shrank over two decades

Recognition never caught up

After more than two decades of paid search, 51% of UK search users can correctly say why the top sponsored results are there, against 54% in 2023 and 48% in 2018. Among 16 to 24 year olds the figure is 37%.¹⁴

Attributed returns overstated true returns

Field experiments at eBay found paid search returns at a fraction of conventional non-experimental estimates. Brand-keyword ads showed no measurable short-term benefit, and heavy repeat buyers who would have purchased anyway absorbed most of the spend.¹⁵ A randomized experiment on a restaurant search app found the reverse error. Labeled ads raised calls to advertisers mainly through exposure rather than clicks. Click-based attribution would have under-credited those ads, and last-click reporting would have assigned much of their effect to organic listings.⁶⁵

4.2 Money

2030 estimates run from $5 billion to $100 billion

Forecasts for assistant advertising disagree by a factor of nearly twenty. Most of the gap is scope. The low figure covers US standalone chatbots, while the high figures are global and include either AI search placements or one company’s own plan.

The largest market-wide number mostly reclassifies existing search spend. More than 80% of EMARKETER’s $32.03 billion for US “AI advertising” in 2026 is ordinary search advertising shown next to an AI-generated summary. Its estimate for ads inside standalone chatbots is under $1 billion for 2026.⁶⁶ OpenAI told investors it expected $2.5 billion in ad revenue in 2026. After the $1 billion run rate at the end of August, EMARKETER’s principal analyst called that target “all but impossible.”^68, 70

ChatGPT ads reached $1 billion faster than search or social ads did

A run rate annualizes the latest period, so it overstates a first year of revenue. OpenAI also launched into an audience of roughly a billion weekly users with self-serve tools, an API and agency relationships that did not exist in 2000 or 2007. Google’s revenue grew fivefold in 2002, the same year AdWords moved to click-based pricing that factored click-through into ad rank, and ChatGPT made the same move from CPM to CPC bidding within three months of launch.^22, 72

4.3 Agents

Platform badges outweighed ad labels in agent simulations

The payment rails for agents buying on a person’s behalf are built. OpenAI’s Agentic Commerce Protocol with Stripe, Google’s Agent Payments Protocol and Universal Commerce Protocol, Visa Intelligent Commerce and Mastercard Agent Pay all launched between April 2025 and January 2026, and Mastercard reported the first agent-initiated transaction on its network in the third quarter of 2025.^{44, 80, 81, 82}

Purchase volume has lagged behind the infrastructure. In March 2026 OpenAI scaled back Instant Checkout inside ChatGPT. Etsy described the resulting sales volume as small, and Walmart saw conversion inside the chatbot run three times lower than when shoppers clicked through to its own site. OpenAI shifted toward retailer apps inside ChatGPT and merchants’ own checkout.⁸³ In US surveys, 6% to 11% of adults report having completed a purchase through an AI tool, and about one in ten would let an agent complete a purchase on its own.⁸⁴

If those patterns hold outside a sandbox, an ad label is a mild penalty when an agent chooses, and influence shifts to platform endorsements, product data and how quickly a merchant responds. Economists studying agent markets expect agents to resist advertising that carries no information or signaling value.⁸⁶ Publishers are already experimenting with ads written for AI crawlers, with no measurement standard yet in place.⁸⁷

A.1 Sources

Where every number comes from

Numbered references appear inline throughout. Each entry gives the date, the kind of source, and, for research, whether it has been peer reviewed. Many of the 2026 academic sources are preprints, and several platform figures reach the public only through trade press relaying company statements.