The Evolution of Online Advertising: From Banners to AI-Generated Responses

Advertising is likely to change more in the next two years than it did in the last twenty. Large language models have moved from tools that help make ads to platforms that serve them, and the rollout is already running. OpenAI began testing ads in ChatGPT in January 2026. Google is placing ads inside AI Overviews. Perplexity has run sponsored follow-up questions since late 2024. The banner era of 1994 to 2000 is the right place to start, because it shows what happens when a format wears out. Banners went from a novelty people clicked to something the entire web learned to ignore, and every model since has been an answer to that collapse.

Digital advertising started on October 27, 1994, when AT&T paid HotWired $30,000 to display a rectangular graphic asking “Have you ever clicked your mouse right here? You will.” That first banner ad achieved a 44% click-through rate according to Digiday’s oral history. Today’s display ads average closer to 0.06%.

The banner era established core patterns still visible in LLM advertising. Publishers needed revenue to keep content free. Advertisers wanted eyeballs. The solution was carving out designated space for commercial messages. By 1995, Yahoo was charging up to $100 for banner placements, and DoubleClick launched tools to track impressions and clicks.

The model worked until users learned to ignore banners. Click-through rates crashed. Pop-up ads emerged as a desperate attempt to recapture attention. By the early 2000s, browsers shipped with pop-up blockers by default. The industry needed a new approach.

The lesson for LLM ads: Banner ads failed because they interrupted rather than assisted. The platforms that figure out how to integrate commercial messages into AI responses without degrading utility will win. The ones that do not will see users flee to ad-free alternatives.

For twenty years, advertising ran on intent. A person typed what they wanted, and the auction matched a commercial message to that request, which made ads useful rather than intrusive. The economics followed. Search grew into a business worth more than $200 billion a year in ad revenue for Google alone, and the industry reorganised around keywords, clicks and cost per acquisition. The mechanism underneath was retrieval. Search engines rank links and hand back a list, so a paid link sits beside an unpaid one and anyone can tell them apart. Language models do not rank and hand back. They write the answer, which removes the slot the entire search model was built to sell.

Search advertising solved the interruption problem by matching ads to intent. When someone searches for “running shoes,” showing running shoe ads is helpful, not intrusive. GoTo.com pioneered pay-per-click in 1999. Google AdWords launched in 2000 and introduced the Quality Score model that rewarded relevance, as documented in HubSpot’s history of online advertising.

The genius of search advertising was alignment. Users wanted answers. Advertisers wanted customers. Google wanted revenue. When the system worked, everyone benefited. That alignment drove over $200 billion in annual ad revenue for Google by the 2020s.

Search advertising worked because it retrieved relevant links. But LLMs do not retrieve. They generate. And that changes everything about how advertising can function.

Three platforms are running live LLM advertising or testing it: Perplexity, Google and OpenAI. Each picked a different place to put the ad. Perplexity keeps it outside the answer, as a sponsored follow-up question. Google puts it inside the AI Overview itself. OpenAI attaches it to the bottom of the response. The approaches differ, the pressure does not. Usage is growing far faster than subscription revenue can cover, so each of them has to monetise a generated answer without spending the trust that makes the answer worth reading. Nobody has solved that, and the formats in market are best read as experiments rather than settled products.

The stakes are enormous. ChatGPT traffic grew 8,400% in a single quarter according to Singular’s analysis. Perplexity processes over 30 million queries daily. These platforms need sustainable revenue models, and subscriptions alone are not cutting it. OpenAI’s CFO confirmed the company explored ad models because subscriptions do not generate enough revenue for long-term sustainability.

Perplexity sells the question, not the answer. The format launched in November 2024 as sponsored follow-up questions. After the AI replies, the user sees a list of related questions to ask next, and some of those slots are paid and labelled as such. Click one and Perplexity's own system writes the response. The advertiser never supplies the copy. That design keeps commercial money one step away from the generated text, which protects the answer and limits how far an advertiser can shape it. It also prices high, because a user who chooses the next question has already declared what they want. Whether that separation survives contact with revenue targets is still open.

According to Perplexity’s announcement, 40% of users click on related questions. The company’s VP of business development described the format as “additive” rather than disruptive. The ad suggests a relevant next question. The AI delivers an objective answer.

Initial partners included Whole Foods, Indeed, Universal McCann, and PMG. Perplexity charges on a CPM basis with rates exceeding $50 per thousand impressions, a premium price point compared to the $2 to $5 average for standard display ads, reflecting the higher intent of the user. By October 2025, Perplexity paused accepting new advertisers to focus on product development, signaling the format is still evolving.

Pro tip: Perplexity’s format keeps ads separate from answers. This preserves trust but limits integration. Watch how user engagement with sponsored questions evolves. It will signal whether separation or integration wins.

Google put the ad inside the answer. Where Perplexity keeps sponsorship in a separate slot, Google places commercial results within the AI-generated summary and matches them against both the query and the text of that summary. Coverage moved quickly once the decision was made, from a small share of keywords at the start of 2025 to a large share by November. Advertisers gain the reach and lose the controls. There is no way to bid for an AI Overview or AI Mode placement specifically, no way to write the surrounding sentence, and no visibility into the matching logic. Google owns all three. Distribution in exchange for control is the defining trade of advertising on a platform that writes the page.

Google took a different path. At Google Marketing Live 2025, the company announced ads would appear within AI Overviews on both mobile and desktop. These ads are integrated directly into the AI-generated summary itself.

According to Semrush’s analysis of 10 million keywords, ads alongside AI Overviews rose from about 3% in January 2025 to roughly 40% by November. The system considers both the user query and the content of the AI Overview when deciding which ads to serve.

Google also began testing ads in AI Mode, its conversational search experience that competes directly with ChatGPT. Ads appear below and integrated into AI Mode responses. The company claims advertisers using AI Max for Search campaigns see 27% more conversions at similar cost-per-acquisition compared to traditional keyword campaigns.

The catch: advertisers cannot yet bid specifically for AI Overview or AI Mode placements. Ads must be highly relevant to both the query and the AI-generated answer. Google controls the matching algorithm entirely.

ChatGPT now carries ads for users who do not pay. They appear under the answer, drawn from the conversation, and only where a sponsored product or service is relevant to it. Paid tiers stay clean. OpenAI is testing this in the United States on the free and Go tiers. The published commitments are specific: ads do not influence answers, conversations are not shared with advertisers, and personalisation can be switched off. The arithmetic behind the reversal is blunt. Roughly 800 million people use ChatGPT each week and about 20 million pay, which is under 3% conversion. Subscriptions alone cannot cover that gap, and the revenue targets attached to advertising run into billions.

OpenAI resisted advertising longer than competitors. CEO Sam Altman previously called ads a “last resort.” That changed in January 2026. OpenAI announced it would begin testing ads for free and Go tier users in the United States.

The company laid out explicit principles:

• Answer independence: Ads do not influence responses. Answers are optimized for helpfulness, not advertiser preferences.
• Conversation privacy: OpenAI keeps conversations private from advertisers and does not sell user data.
• Choice and control: Users can turn off personalization and clear data used for ads. Paid tiers remain ad-free.

The initial format shows ads at the bottom of answers when there is a relevant sponsored product or service based on the current conversation. Internal documents reported by Digiday show OpenAI planning $1 billion in “free user monetization” starting in 2026, scaling to nearly $25 billion by 2029.

The challenge is enormous. ChatGPT has approximately 800 million weekly users, but only about 20 million pay for premium tiers. That is under 3% conversion. Advertising economics become essential at that scale.

The next auction may not sell space at all. It may sell wording. In a token auction, advertisers submit a fine-tuned model carrying their brand voice instead of ad copy. They then bid on how strongly that model pulls each word the system generates. Buy the adjective, not the slot. This is not a thought experiment left in a drawer. It has been tested on a small open model, higher bids produced measurably stronger influence over tone, and Google filed a related patent as far back as 2018. If it ships, the labelling problem gets much harder. There is no ad to point at, only a sentence that leans.

Here is where it gets interesting. Google Research and the University of Chicago published a paper outlining a theoretical framework for token auctions. Instead of bidding for ad slots, advertisers would bid to influence the actual words an LLM generates, token by token.

In this model, advertisers do not submit static ad copy. They provide their own fine-tuned language models representing their brand’s voice, tone, and messaging. When a user enters a query, the system generates a response one token at a time, weighing how much each advertiser’s model should influence the next word based on bid strength.

Think of this as bidding for adjectives. In a standard auction, a hotel chain bids to show their link when someone searches “best hotels.” In a token auction, that same hotel chain might bid to increase the mathematical probability that the word “luxury” appears next to their brand name in the generated answer. The brand buys a subtle shift in the AI’s vocabulary, not a slot.

The researchers tested this using Gemma 7B with two dummy advertisers. Results showed clear correlation between higher bids and stronger influence over tone and wording. One advertiser bidding 3x more than another shifted the generated text predictably toward that brand’s voice. The full paper is available on arXiv.

This is not science fiction. The paper won the WWW 2024 Best Paper Award. Google filed a related patent in 2018 titled “Using various AI entities as advertising mediums.” The infrastructure for this future is being built now.

The trust risk: Token auctions could blur the line between information and advertisement. If users cannot tell whether a response is influenced by advertiser bids, trust erodes. The platforms that maintain transparency will likely win long-term, even if they sacrifice short-term revenue.

There is no self-serve LLM ad platform to buy on yet, which makes preparation the only available move. Four things are worth doing now. Make the brand easy for models to find and cite, because the content these systems quote is the content they end up recommending. Shift targeting thinking from keyword lists to conversation context, since the match happens on the problem rather than the phrase. Decide in advance what the model is never allowed to say about the product, and ask platforms for that constraint in writing. Set aside a budget for the beta window. The advertising model that paid for free search for 25 years is being replaced.

Make your brand LLM-findable. The content AI systems cite becomes the content they recommend. Structure your site with clear, answer-first content that LLMs can easily parse. Schema markup, clear headlines, and factual claims with sources increase citation likelihood. Search Engine Land’s GEO coverage tracks how this field is evolving.

Prepare for context-based targeting. Keyword lists will not transfer directly to LLM advertising. The systems match based on conversation context, not just query terms. Think about the problems your customers are trying to solve, not just the words they might type.

Watch the pioneers. Perplexity’s sponsored questions, Google’s AI Overview integration, and ChatGPT’s bottom-of-answer placement are experiments. Track what is working, what is getting backlash, and what is evolving.

Define your hallucination tolerance. In traditional search, you control the ad copy. In LLM advertising, the model generates the copy. What happens when a sponsored answer promises a feature your product does not have? Marketers need to establish strict brand safety guidelines and demand “negative constraint” capabilities from platforms. These are lists of things the AI is never allowed to say about the brand.

Budget for experimentation. When self-serve LLM ad platforms launch, early movers will have advantages. Lower CPMs during beta phases. Algorithm influence from early campaign data. Category ownership before competitors arrive.

The advertising model that funded free search for 25 years is evolving. The brands that understand where it is heading will shape the next era. The ones that do not will pay premium prices to catch up.