AI Agents, Search Ads and the CPM Paradox: A Deep Dive

The CPM paradox – why it matters now

The growing conflict between AI agents and search ads is challenging the very foundation of search advertising that underpins the free web. In 2024 the U.S. search-advertising market generated $102.9 billion and made up almost 40% of all digital ad revenue. The business still rests on a simple idea: advertisers pay for a thousand impressions (CPM) because a person might see and click. But generative AI has inserted itself between consumers and the web. Large-language-model (LLM) releases such as GPT-4o, Claude Sonnet and Gemini 2.0 now act as autonomous assistants that synthesise answers instead of listing links. They plan trips, interpret code and can complete purchases. They also “see” search ads and content long before people do.

The misalignment is profound. While algorithms filter and summarise information, search advertisers still pay CPM rates as if human eyeballs were watching. This is the CPM paradox: paying for impressions when machines are the ones viewing. It raises fundamental questions about the value of exposure and future of search advertising.

How AI agents view ads – and why CTRs are falling

A research team at the University of Applied Sciences Upper Austria simulated hotel-booking tasks for GPT-4o, Claude Sonnet 3.7 and Gemini 2.0 Flash. They found that GPT-4o selected a single hotel in 95% of trials; Gemini chose one only 60% of the time. GPT-4o completed 84 bookings versus 43 for Gemini. The agents used keywords and structured data to filter options; visual banners were largely ignored. In other words, persuasive copy and imagery—the bread and butter of creative advertising—often never “exist” for an AI. If your product feed lacks prices, availability and attributes, an assistant may simply skip it.

AI Summaries and the Collapse of CTR

Consumer behaviour is also changing. Google’s AI Overviews, which summarise search results, appeared in about 6.49% of queries in January 2025 and 13.14% by March 2025. Click-through rates (CTR) fell sharply too. One study found organic CTR dropped from 1.76% in June 2024 to 0.61% by September 2025, while paid CTR fell from 19.7% to 6.34%.

Another Digiday analysis noted that AI Overviews drove zero-click searches up to nearly 69% and plateaued at roughly 20% of desktop search queries. Pew Research reported that when an AI summary appeared, only 8% of users clicked a result link versus 15% on pages without a summary; just 1% clicked a link inside the summary. Over 26% of users ended their session after reading the summary, indicating that the summary itself satisfies intent. Chatbot adoption is still growing too. Only around 3% of search-related traffic is handled by chatbots—yet the effect on click-throughs is already material.

AI search interfaces are also changing the way people articulate queries. Google’s experimental AI Mode—designed for complex, conversational tasks—encourages users to ask much longer questions: the company notes that queries in AI Mode are nearly twice as long as traditional searches. The longer query length provides more context for targeting but further diminishes the value of impression-based pricing, as the assistant synthesises an entire answer rather than listing links.

The monetisation dilemma: CPM, CPC and the quest for outcomes

Despite machines mediating the experience, many platforms are struggling to price AI agents and search ads, often defaulting to selling them as though humans are watching.. Perplexity AI, a Gen-AI search engine, launched a CPM-based model charging US $30–$60 per thousand impressions, roughly eight times premium display rates. Marketers told Digiday that these CPMs were “totally out of whack” and risked paying for unmeasurable impressions. Another analyst said pricing AI search on impressions is “risky” because there is no way to verify performance. South Korea’s Liner, by contrast, offers a cost-per-click (CPC) model with a click-through rate of 1–2%, aligning cost with measurable actions.

Emerging models compared

Google and Microsoft are experimenting within their ecosystems. Google’s AI Overviews insert sponsored links into answers and now account for a growing portion of queries. Microsoft’s Copilot surfaces ads inside chat modules. Yet neither company has disclosed pricing, and the declining CTRs noted above suggest that impression-based pricing may not reflect value. Pew Research found that when AI summaries appear, fewer users click through, further eroding the CPM value proposition.

Agencies push back – and call for outcome-based metrics

Advertisers are understandably sceptical of paying premium CPMs for machine-mediated “impressions.” In Digiday’s reporting, executives likened Perplexity’s CPMs to Netflix’s premium CTV rates and said the terms “reek of desperation.” Ryan Bopp, then senior vice-president of digital strategy at the Eden Collective (Bleacher Report’s parent company), contrasted Perplexity’s rates with TV advertising: he noted that a 30-second screen takeover on connected TV costs about the same as a display or sponsored prompt on Perplexity, which he called “disproportionate.”

Another search specialist warned that a CPM-based model makes little sense for search because advertisers expect a click-based offering. Debra Aho Williamson,stressed that the most successful digital businesses (like Google and Meta) are built on performance advertising. She cautioned that companies launching with premium brand-awareness models often struggle. Liner’s CPC model feels more familiar because search advertisers are accustomed to bidding on clicks.

Outcome-based pricing is gaining momentum. A sponsored post on AdExchanger argues that CPMs “penalise efficiency and experimentation” and that ad-tech partners should move to unlimited, outcome-based pricing, aligning cost with results. In other words, advertisers should pay when an agent actually books a hotel, adds a product to the cart or completes a sale—not when a bot “sees” the ad. However, verification remains a challenge: there is no standard method to measure when an AI agent reads an ad or picks a product. The tension between advertisers and platforms underscores the need for new metrics that capture agent-level actions rather than impressions.

Creative and strategy: optimising for machine logic

As AI assistants filter content, the creative strategy for AI agents and search ads must evolve.. The Austrian hotel-booking study showed that agents make decisions based on structured data and keywords, not emotional imagery. To appeal to machines, brands need to supply machine-readable content—clear prices, specifications, availability and reviews—and embed keywords in visible text. Michael Lehman’s “Marketing to Machines” column on AdExchanger notes that LLM-powered agents act as customers: they read, compare and synthesise rather than clicking or scrolling. The goal, he argues, is no longer to rank number one on a blue-link SERP but to become part of the answer. This requires Generative Engine Optimisation (GEO): narrative authority, factual depth and structured data that an AI can trust.

Practical steps include expanding product feeds, using schema markup, providing descriptive alt text and ensuring Q&A-style copy so AI systems can extract concise answers. Such optimisation blurs the line between marketing and product data; brands must treat AI agents as their first audience.

What’s next: agentic commerce and browser wars

The future of search advertising may look less like a list of links and more like agentic commerce, where assistants not only answer questions but execute purchases. OpenAI’s ChatGPT recently introduced Instant Checkout: U.S. users can buy items from Etsy and (soon) from more than a million Shopify merchants directly inside ChatGPT. Merchants pay a fee for completed purchases, and OpenAI plans to open-source the protocol built with Stripe.

Amazon’s Rufus AI assistant is moving in the same direction. According to Modern Retail, Amazon quietly added an Auto Buy feature that monitors prices and automatically purchases products when they hit a shopper’s target price; users have 24 hours to cancel. Rufus users were 60% more likely to complete a purchase and Amazon expects the assistant to generate over US $10 billion in incremental annualised sales. Hindustan Times notes that the feature allows the AI to track price drops and automatically place orders.

New AI-first browsers are also emerging. We have the likes of Comet, Perplexity’s AI-powered browser that can summarise emails, browse websites and manage tasks. Other companies have launched AI browsers as well; Dia from The Browser Company, which features an AI chat tool; Opera’s Neon, which works offline; and OpenAI’s Atlas that operates in “agent mode”. These browsers compete to become the primary interface between users and the web, inserting AI at every step.

Adoption and the scale of the shift

Despite the hype, the shift to AI-mediated search is still early. A BrightEdge report found that AI search drives less than 1% of referral traffic. This shows that organic search remains the primary channel for conversions. Birdeye notes that chatbots handle only around 3% of search-related traffic and that Google still controls about 89.6% of global search. Even ChatGPT’s scale is modest relative to the incumbent: the platform processes about 2.5 billion prompts per day, while Google handles roughly 16.4 billion searches daily, and only 61% of Americans have used any AI tool in the past six months. Pew Research found that only 18% of Google queries in March 2025 produced an AI summary.

Yet the momentum is real. McKinsey reports that about half of consumers already use AI-powered search and projects that AI-mediated search could influence US $750 billion in consumer spending by 2028.

The consultancy expects AI summaries to appear on three-quarters of search queries by the end of the decade and notes that 44% of AI search users consider answer engines their primary source of buying decisions. In other words, while generative search today accounts for a small share of traffic, the trajectory points toward mainstream adoption within a few years.

Headwinds: why AI search is still nascent

AI’s growing role in commerce is constrained by both technical and regulatory headwinds. Generative models remain unreliable narrators: a recent benchmark found that even the best models produce hallucination-free text only about 35% of the time, leading a Cornell researcher to note that we cannot yet fully trust AI outputs.

The same study also found hallucination rates vary by topic, with some models refusing to answer. Prominent mistakes, like Google Bard’s $100 billion error, damage credibility. These issues highlight why prominent mistakes—such as Google’s Bard chatbot sharing a factual error that wiped $100 billion off Alphabet’s market value—damage credibility. Further limitations include constrained context windows and difficulty accessing real-time data, which make it hard for assistants to remember personal preferences or fetch up-to-date prices. For now, most AI assistants act as research tools rather than conversion engines. BrightEdge reports that AI search drives less than 1% of referral traffic and Birdeye notes that chatbots handle only around 3% of search-related traffic. In other words, adoption is growing but still nascent.

Legal and regulatory scrutiny compounds these technical challenges. The U.S. Federal Trade Commission’s Operation AI Comply argeted companies making false claims about AI capabilities. This action demonstrates regulators’ willingness to police deceptive AI marketing. The European Union’s AI Act requires AI systems that generate synthetic content to include machine-readable labels indicating that the content is AI-generated. These rules will force platforms to clearly disclose AI-authored answers and ads. Pew’s survey shows that many users are less likely to click links within AI summaries, underlining the importance of transparency and accuracy as AI becomes the interface for commerce.

Conclusion and action plan for marketers

The CPM paradox shows the core misalignment between AI agents and search ads. Platforms are still charging for human views, even when machines are the ones doing the viewing.

This misalignment has several effects. AI agents evaluate structured data, not banners. Click-through rates (CTRs) are declining because AI summaries satisfy user intent. Early monetisation experiments also reveal friction between old CPM pricing and new outcome-driven expectations

Key takeaways:

1. Treat AI agents as your first customers. Build rich product feeds with prices, specs, availability and user reviews. Machine-readable data—not glossy imagery—is what assistants parse.
2. Optimise for machine comprehension. Use schema markup, descriptive alt text and Q&A-style copy to ensure AI assistants can extract accurate information. Embrace Generative Engine Optimisation to become part of the answer, not merely the top link.
3. Push for outcome-based deals and new metrics. Advocate for pricing tied to bookings or conversions rather than impressions. Demand agent-level logs and verification mechanisms.
4. Experiment with answer-engine optimisation and agentic commerce. Consider plug-ins, APIs and AI-native storefronts (e.g., ChatGPT’s Instant Checkout or Amazon’s Rufus). Monitor how your brand appears in AI answers and adjust feeds accordingly.
5. Keep perspective. AI search is growing fast but still small; organic and paid search continue to drive most conversions. Maintain core search budgets while testing new channels, and demand transparency as technology and regulation evolve.

The winners of the next era of search will not be those shouting for attention but those whispering the right data into the ears of machines. By aligning creative, pricing and strategy with machine logic, marketers can turn the CPM paradox into a catalyst for innovation.