Ads won't be the backbone of the AI economy

Build an audience. Subsidise the experience. Sell access to the audience.

AI has every reason to reach for the internet’s familiar business model. Call it commercial reversion to the mean. Advertising is available, understood and capable of funding free access. I expect it to remain part of the mix.

But I doubt it becomes the backbone of the AI economy.

Facebook was selling ads in 2004, before launching its broader Facebook Ads system on 6 November 2007. Its eventual model was a commercial choice, not a law of the internet.

AI platforms are making different choices already. OpenAI began testing ChatGPT ads on 9 February 2026 and announced a new visual format on 5 October. Anthropic committed to keeping Claude ad-free on 4 February. Perplexity abandoned advertising, citing trust concerns, as the Financial Times reported in February.

The trust problem

An assistant is supposed to give me the best answer it can. Advertising introduces another customer’s objective into that relationship.

OpenAI says ads do not influence its answers and are displayed separately. That distinction matters. My concern is what happens as revenue targets grow: the incentive to sell influence remains, even when the interface draws a line around it.

For a product whose value depends on factual reliability, I think advertising is the least attractive foundation.

Where I see the money

I’ve argued that the economics of AI remain unsettled. Putting aside the worsening value exchange with content originators that motivates my work on Odexa, three opportunities look more durable.

Content and citation agreements. Bilateral deals can price access, reuse and attribution for valuable content and data. These create revenue for originators and costs for platforms, which must recover those costs through useful services. Paying for access must never become paying to be cited as correct.

Insight from the source. With permission, platforms could turn aggregated demand signals into useful business intelligence: what customers need, what confuses them and where products disappoint. Businesses would pay for better decisions. That proposition depends on protecting the people behind the data.

Agentic commerce. When an agent completes a booking, purchase or procurement task, there is a concrete outcome against which to charge a fee. OpenAI’s September 2025 Instant Checkout launch included merchant fees on completed purchases. Its March 2026 shift towards merchant checkout experiences also shows how unsettled the execution remains.

Commissions have conflicts too. An agent must disclose fees and choose for the customer, even when another seller pays more. My bet depends on that trust being earned.

Open models change the calculation

As of October 2026, I am less convinced that selling access to proprietary models alone provides a durable business. Mozilla’s September assessment found eight of OpenRouter’s ten leading models by August token volume had open weights. Its developer survey also showed substantial coexistence: half used both open and closed models.

That is meaningful competitive pressure, with limits: OpenRouter represents one platform, and developer adoption is not enterprise spending. I see pressure on what a subscription must deliver. Hosting, integration and dependable operation can still command a price, even when the underlying model is open.

I expect value to accumulate around trusted information and useful work completed. Agentic commerce gives platforms a way to participate in that value, with fees tied to transactions people actually want.

Advertising will help pay some bills. Despite its familiarity, I don’t expect it to form the backbone of the AI economy.