What are AI walled gardens?

What makes a garden walled

Not size, and not being proprietary. A walled garden is defined by the same party controlling all four of: what inventory exists, how it is bought, how performance is measured, and who the audience is. When one company holds all four, an advertiser cannot independently verify what they received, cannot compare it against another surface on equal terms, and cannot take their audience elsewhere.

The AI surfaces have every structural advantage the previous generation had, and one more. Their inventory is not just theirs — it is generated by their model, in their voice, inside a product with no external equivalent. There is no arbitrage, no reseller, and no way to reach that user in that moment anywhere else.

What it costs a buyer

• Marked-your-own-homework measurement. Performance is reported by the party being paid, with no independent counting layer.
• No cross-surface comparison. Metrics are defined differently on purpose, so like-for-like evaluation is not possible.
• No portability. Audience understanding accumulated inside the garden does not leave it.
• Terms set unilaterally. Pricing, format and policy change on the platform's schedule, not on a contract's.

The counterweight

The alternative is not to refuse to buy the large surfaces — that would be self-defeating while the attention is there. It is to keep a second, independent path viable: supply that can be bought on open protocols, measurement that exists outside any single platform, and reach across the long tail of AI publishers whose inventory is often more intentional than a general assistant's.

That is the case for the AI open web and for standards such as AdCP. It is an argument about market structure rather than about product quality, and it is made most loudly by the parties who lose if the gardens win — which does not make it wrong.

What a buyer can do now

Three habits keep options open at almost no cost. Run an independent holdout so you have a performance read the platform did not produce. Keep at least some spend on independent supply so the capability exists when you need it. And keep creative and product data in a portable structured form rather than in one platform's format.