Is your protocol on the list AI agents are already trading from?

Changpeng Zhao, known throughout crypto as CZ, posted on X that AI agents will make one million times more payments than humans, using crypto to do it. The shift he described is not theoretical. It is already happening.

By early 2026, x402 had already powered more than 100 million agentic transactions on Base, according to on-chain analysis from Chainalysis. A large share of that early volume came from a single memecoin "pay-to-mint" project rather than organic agent commerce, but the underlying pattern held: these were transactions built for software, not people. No browser needed to be opened. No search needed to be run. An agent could decide where the money went and send it.

This changes something specific for protocols. Not traffic or rankings. It changes the identity of the buyer and who makes the decisions. Appealing to human buyers is no longer enough. Protocols must now be discoverable, understandable, and trusted by the software agents increasingly deciding where transactions go.

Why did crypto protocols never bother with Google SEO?

Ask most protocol founders or marketing leads about SEO and the reaction is usually the same: a shrug, a grimace, maybe a story about the agency they hired and fired three months later. For most of crypto's history, that reaction made perfect sense.

Look at how the biggest protocols actually grew. Uniswap did not rank its way to a user base with SEO. It airdropped UNI to people who had already used the protocol, and attention followed the money. NFT projects built communities, demand, and entire economies inside Discord servers before many buyers had ever visited a website. Liquidity mining programmes paid users to show up, use the product, and bring others with them.

Protocols didn't grow through blogging, guest posts, or buying links through an agency. The growth strategy was to go where the market already was: Discord, Telegram, X, crypto Twitter, token incentives, partnerships, KOLs, community. Google sat much further down the list.

There was a good reason for that. The audience was not searching and comparing websites on Google. They were hearing about a protocol from someone they trusted in a Telegram group, seeing it discussed on X, checking the Discord, reading the docs, and forming an opinion from there.

Building a large SEO content operation for an audience that rarely used search to make decisions would have been wasted spend.

Spend enough time actually inside these servers and the pattern becomes obvious. The "marketing" was never the marketing team. It was people who genuinely believed in the thing, repeating the pitch in their own words, faster and more convincingly than any blog article or marketing guru ever could.

That mechanism worked because it had one assumption baked into it: a human being was always the one reading the group chat, doing the research, and forming the opinion. That underlying assumption no longer holds.

The "but nobody's searching for us" objection we hear from founders and marketing leads is entirely valid on its own terms, but it was always about traffic: humans typing queries into Google, comparing pages, clicking a result. None of that happens here. No page gets visited. No ranking gets checked. An agent asks a model a question, the model answers from whatever it already trusts, and money moves. Being invisible on Google search was survivable when Google was optional. Being invisible to the model itself isn't, because the model is no longer offering a list of links for a human to sift through. It's making the call directly.

What's actually happening with AI agents making crypto transactions right now?

CZ's prediction wasn't posted in a vacuum. Coinbase built x402 specifically so software agents could hold funds and pay for things without a human approving each transaction. It has since moved to the Linux Foundation, backed by Visa, Stripe, AWS, and Google Cloud. It's no longer one company's side project. It's shared, neutral infrastructure.

Within hours of CZ's post, Coinbase CEO Brian Armstrong made a related, practical argument: AI agents cannot pass Know Your Customer checks, so they cannot open a bank account. But they can hold a crypto wallet with nothing more than a private key. No forms. No approval process. No human in the loop at all.

Coinbase had already put that argument into a product before either CEO said a word about it publicly. Agentic Wallets launched in February 2026, built for agents from the ground up rather than retrofitting a product originally designed to be human-facing. Two months later, Coinbase opened Agent.market, a marketplace organised entirely around bots buying and selling from each other across categories like data, trading, and infrastructure, settled in USDC. Other platforms are building the same idea from a different angle: Virtuals Protocol's Agent Commerce Protocol handles requests, negotiations, transactions, and evaluations between AI agents directly, and BNB Chain's ERC-8004 and BAP-578 standards give software agents verifiable on-chain identities and the ability to hold and spend funds without human authorisation at the transaction level. This isn't a single company's bet. Several chains are building toward the same outcome.

None of this required a person to search for anything on Google. An agent gets instructions, holds a wallet, and pays. The decision about who gets paid happens entirely inside a system most protocol teams have never looked at.

This doesn't mean handing control of a wallet completely over to an AI agent. EIP-7702 and session keys let an agent act inside narrow, temporary limits: a fixed spending cap, a specific set of allowed actions, an expiry time. The person who owns the wallet keeps full custody while giving the agent just enough authority to do its job and nothing more.

This is where the problem for protocol marketers starts. Everything that built your user base until now was built for a reader. AI agents don't care whether your Discord is active. They don't scroll group chats, don't read pinned messages, and don't pick up on community sentiment. They ask a model a question and act on that answer.

That matters because an AI agent cannot be influenced by a conversation it never sees. If the signals that make people trust your protocol exist only inside closed communities, they may as well not exist at all to the agent deciding where the money goes.

Your protocol can be trusted by thousands of people and still be invisible to the machine now making the decision.

How does an agent actually decide who to pay?

There are two separate decisions here, and conflating them is where many explanations of this topic go wrong.

The first is mechanical. Once an agent knows what it wants, "swap 1,000 USDC for ETH at the best rate," it doesn't pick a protocol by reading a marketing page or browsing a Discord server. It states the outcome, this is what's called intent-based execution, and a network of solvers competes to fill it, each one racing to offer the best price and route. Tools like 1inch, CoW Swap, or Jupiter aggregate across dozens of liquidity pools automatically. This layer runs on pure execution quality: price, slippage, gas cost. A protocol can't market or SEO its way into a better swap rate. This part of the system genuinely doesn't care about your content or marketing strategy.

Almost all of this settles in stablecoins, not volatile tokens. Of the agent-to-agent payments already tracked on-chain, over 98% have gone through USDC. Stablecoins are the default rail here for the same reason they're the default rail everywhere else in crypto: an agent doing math on a swap or a yield strategy needs a predictable unit of account, not something that might move 5% before the transaction settles.

The second decision happens earlier, and this is the one that actually matters for protocols.

Before an agent can state a precise outcome, something has to narrow the field. "Swap USDC for ETH" is precise. "Find the safest yield strategy" is not, and a huge share of what agents are actually asked to do looks like the second kind, not the first.

That ambiguity has to be resolved somewhere.

When the instruction is vague, the agent has two choices: ask a clarifying question, or lean on the model behind it to fill the gap, which protocols exist, which ones are considered reputable, which one fits "safest." That's not a solver auction. That's the model reaching for whatever it already knows or can retrieve.

Models are not especially cautious when they do this. A recent academic benchmark testing agents on ambiguous instructions across several domains, including a trading bot, found that standard agents reached the correct outcome only 43% of the time without assistance. Left to resolve ambiguity themselves, they usually didn't stop and ask. They picked.

That pick has an economic consequence for protocols. Whatever gets picked isn't just an answer. The protocol the model reaches for is the protocol that gets considered. The protocol that gets considered has a chance to be routed. And the protocol that gets routed earns the fee.

Multiply that selection moment across millions of ambiguous agent instructions and protocol discovery stops being a marketing nice-to-have. It becomes a distribution layer for transaction flow.

So, for a precise, already-decided transaction, you're invisible to the process, and that's fine: nobody markets their way into a better swap rate. But for anything with ambiguity in it, and most real instructions have some, you're back to being one of the sources the model either knows about or doesn't.

For protocol founders and marketing leads, that's the uncomfortable part: your next competitor may not beat you on APY, liquidity, gas fees, or execution. They may simply be the protocol the model already knows to reach for when nobody's told it what to choose.

So what does the model reach for when it has to guess?

The previous section established the mechanism: when an agent's instruction is ambiguous, the model behind it reaches for whatever it already knows or can find. That's not a hypothetical. It's the exact thing our own research already measured, just for a different audience.

A 50-protocol audit across ChatGPT, Perplexity, and Google AI Overviews, over 1,000 citation records, found that official protocol pages made up just 1% of what Perplexity explicitly attributed claims to, and 4% of what Google's AI Overviews attributed claims to, when answering questions about crypto protocols. The study distinguishes this from simply being named in an answer: a citation means the system pointed at a specific source for a specific claim, not just that the protocol's name came up somewhere in the response. Instead, these systems cite YouTube videos, Reddit threads, and third-party guides. Six protocols in the study scored zero, named nowhere, cited nowhere, on any of the three platforms, for any query.

There's a straightforward reason for this. These systems are built to answer a question the way it was actually asked: "is this safe," "how does this compare," "what happened when this protocol got exploited." Official pages are written to describe a product, not to answer that shape of question. A Reddit thread arguing about which lending protocol is safer, or a YouTube video walking through a hack post-mortem, matches the query far more closely than a homepage ever will, so that's what gets pulled in and cited. It isn't that official pages are penalised. It's that they're rarely built to be the thing a model would reach for.

That finding was built around how humans use AI assistants to research crypto. But the mechanism doesn't change depending on who's asking. An agent resolving "find the safest yield strategy" draws on the same underlying model, trained on the same corpus, weighted toward the same third-party sources, as a person typing the same question into ChatGPT. If your protocol is invisible to one, it's invisible to the other. The only thing that's changed is who's reading the answer.

For protocol founders, this changes where the commercial risk sits. You can have the deepest liquidity, the best execution, and the strongest risk controls in the market, but none of that matters if the model never puts you into the agent's consideration set. And for marketing teams, the problem is worse than a missed click or a lower search ranking. There is no search results page, no homepage visit, and no user to retarget. The agent makes the choice upstream, using sources you may not own and narratives you may not even know are shaping the decision. By the time the transaction reaches the execution layer, the protocol competing for that revenue may already have been chosen.

What does an agent actually see when it looks at your protocol?

Theory is one thing. Here's what it actually looks like.

We ran one query, "find the safest DeFi yield strategy for stablecoins," no protocol named, no platform specified, across ChatGPT, Perplexity, and Google AI Overviews independently. Worth being precise about what this is and isn't: it's not a trace pulled from inside a live trading agent, that infrastructure is proprietary. It's the same proxy method our original study used: asking the models that sit behind these agents the kind of question an agent would be resolving, and recording what comes back.

All three converged on the same answer.

ChatGPT recommended Aave by name, with a live APY figure, alongside Sky's sUSDS, and explicitly excluded Morpho, giving its reasoning: vault risk there depends on curator-set allocations, an extra layer of analysis it wasn't willing to assume on the user's behalf. Perplexity named Aave v3 as a core option too, alongside Compound and Morpho-on-Aave. Google's AI Overview led with Aave and Morpho Blue as its top blue-chip lending pick, plus the Sky Savings Rate.

Nobody asked for Aave. It won by default, three times, independently, simply by being the protocol each model already trusted enough to reach for.

There's a second finding worth sitting with. ChatGPT ruled Morpho out with a stated reason. Perplexity and Google both included it without hesitation. That's not noise, it's texture: even among the protocols models reach for by default, there's real disagreement on the next tier down, which means this isn't a simple popularity contest. Something in what each model has retrieved about Morpho specifically is shaping that split.

The sourcing behind Google's answer makes the earlier point concrete rather than abstract. Its supporting links weren't Aave's site, or Sky's, or anyone's official page. They were a YouTube channel, a Reddit thread, and two third-party finance sites. Not one protocol's own domain made the list.

That's the whole argument in one screenshot. The model didn't need to visit your site to make a decision about your category. It needed something else to have written about you first.

Frequently asked questions

• Do AI agents really hold crypto wallets?

  Yes, already, not as a future concept. Coinbase's Agentic Wallets, built specifically for AI agents rather than human users, launched in February 2026 on the x402 payment protocol, which had processed more than $600 million in transaction volume by early 2026. Session keys and standards like EIP-7702 let an agent act within a fixed spending cap and a limited set of permitted actions, while the wallet's owner keeps full custody. The agent gets enough authority to do its job and nothing more.

• Is this only relevant to trading bots?

  No. Anything with an ambiguous instruction behind it is affected, not just active trading. "Find the safest stablecoin yield" is exactly the kind of query where a model fills the gap with whatever it already trusts. Lending, staking, and treasury management instructions carry the same exposure as swaps and trades.

• What can a protocol actually do about this?

  The lever isn't technical. It's the same finding as our original citation study: the protocols that get named consistently have a distributed record across third-party platforms, developer docs other tools reference, coverage that journalists and forums cite, explanations that show up on aggregators. Working out exactly where those gaps sit for a specific protocol means mapping what's actually being cited across its category and why, not just checking whether the protocol's name comes up. That's the starting point for anyone serious about closing the gap rather than guessing at it.

• Does this replace the need for a normal website?

  No. A model still needs somewhere to confirm what your protocol is, what it does, and how it relates to other named entities in its category. A homepage doesn't earn citations on its own, but it remains part of how a model resolves entity identity once something else has already made you worth citing.

Revenue in this industry has always followed volume. That hasn't changed. What's changing is who decides where the volume goes. For years, that decision belonged to a person: reading a Discord, trusting a group chat, comparing protocols, and eventually clicking approve. Increasingly, it will belong to a model resolving an agent's instruction and reaching for the protocols it knows well enough, and trusts enough, to choose. The transaction may still settle on-chain. The fee may still accrue exactly as it does today. But the decision about who earns it is moving upstream. Protocol teams have spent years competing to be where users are. The next fight is to be present where machines make choices.