AI doesn't answer from the model alone: brand information, product specs, credentials, and reputation all live outside the model and have to be fetched before answering. Split the ecosystem by role and it comes apart into five layers. Here is what each one does, who plays there, and why supplying accurate brand facts is a layer of its own.
2026-08-22
Most writing about AI stares at the models: whose is stronger, when the next generation ships. But between a user’s question and the answer they get sits a lot more than a model. Split the chain by role and you can see who does each part, what is changing, and what isn’t; you also see where a brand actually stands in this ecosystem.
Why split it into layers at all
A concrete case: a user asks AI “which brand in this category is worth buying.” However strong the model, its parameters don’t contain your product’s price this month, the certification you renewed last week, or yesterday’s customer reviews. Those live outside the model. Before answering, the AI has to retrieve them, then decide what to trust, what to use, what to cite.
So a reliable answer passes through at least: the model’s reasoning, the pipes that fetch outside information, the quality of that information itself, and the device where the answer reaches the user. Each link has a different maturity and a different competitive picture. Blur them together and you land on conclusions like “a stronger model will fix everything”, which it won’t.
The five layers
| Layer | What it does | Who’s there | State of play |
|---|---|---|---|
| 01 Models | Training and inference | OpenAI, Anthropic, Google, ByteDance | Extremely high barriers |
| 02 Retrieval | Piping external information into models | Google, Bing, Bocha | Unsettled |
| 03 Fact supply | Feeding AI accurate, structured brand facts | That’s us | No standard set yet |
| 04 Transactions & payments | Agent-to-agent commerce and settlement | Stripe, Visa, UCP | Giants moving in |
| 05 End devices | Where people and agents meet | Apps, hardware, embodied AI | Everyone claiming ground |
01 Models. The source of capability, and of the burn rate. Training a frontier model costs billions, which guarantees very few players. For everyone else, models are infrastructure to call, not a battlefield to enter.
02 Retrieval. The search backends models use for real-time questions. Bindings differ: some use Bing, some build their own; in the Chinese ecosystem, providers like Bocha supply search to multiple models. The pipes decide which sources AI can see at all, and the picture is far from settled.
03 Fact supply. The quality of what the pipes bring back. The problem at this layer: what the internet says about a brand is usually outdated, contradictory, or secondhand. For every AI to cite a brand accurately, its facts have to be structured, consistent, verifiable, and placed on the sources each ecosystem crawls. This is where we work.
04 Transactions & payments. Once agents start ordering on people’s behalf, “how a deal is struck and how money moves” needs new protocol rails: Stripe, Visa, and protocols like UCP are all positioning here. It sits on top of the first three layers: an agent needs trustworthy information before it can transact for anyone.
05 End devices. Apps, hardware, embodied AI: wherever people meet agents. Devices are multiplying, but whichever one wins, it calls the same chain underneath.
Why better models won’t absorb layer three
The obvious objection: as models improve, won’t they swallow the middle layers?
The model and device layers are indeed moving fast. But the need at the fact-supply layer is exactly what stronger models can’t remove, for two reasons.
First, facts are produced outside the model. Your new price, your new certification, your new customer reviews happen in the physical world and in business systems, not in any model’s parameters. Every model generation still needs someone to organize and supply those facts. The models change; the structure (answers depend on external facts) doesn’t.
Second, the ecosystem is fragmented. Every AI crawls different sources, trusts different signals, cites in different ways. What Doubao trusts and what ChatGPT trusts are not the same list, and the Chinese and English ecosystems are two different worlds. For brand facts to be cited accurately across platforms, they have to be supplied to each ecosystem on its own terms. This fragmentation isn’t converging as models improve; in recent years it has been widening.
One line summarizes the division of labor: data companies serve model training; we serve model inference. Training decides what a model knows; inference decides what it says right now, and your customer’s question happens now, so the answer has to be right now.
The value accumulates
One more property worth calling out on its own: some of what models cite at answer time makes its way into the training corpora of later models, gradually becoming knowledge the model simply has. The facts you supply today shape more than today’s answers.
Nobody can promise the rate or the share: training-data selection rules are unpublished and changing. But the direction is clear: this work has an order to it, and brands that start earlier build a deeper base in AI’s understanding. The background numbers are in our AI shift dataset: both content production and web requests are now majority-machine, and the pool that settles grows daily.
What this means for a brand
A brand doesn’t need to join the model race and can’t move the retrieval picture. But the work at layer three directly decides how you appear in AI answers. Concretely, three things:
- Build the fact source. Who you are, why you’re credible, how you differ: all organized into one accurate, consistent, verifiable set of brand facts.
- Make it readable. Your site and content in a form machines can crawl, parse, and cite, placed on the sources each ecosystem draws from.
- Keep verifying. Stress-test on a regular cadence with AI recommendation and fact citation rates, confirming AI gets you right and cites you as the source.
Together, those three are what we call A2A Marketing.
Further reading: The AI shift, in numbers · The AI marketing vendor landscape · About Daimonia