A claim is moving through Silicon Valley this spring: AI agents are the new buyers of the internet, and businesses without an MCP server will be invisible to them. The claim is partly true, which is what makes it dangerous — because the part that is wrong is the part most builders will optimize for.
The argument runs like this. Agents are increasingly the layer through which purchase decisions get made. Restocking, scheduling, booking travel under a price ceiling, picking among interchangeable SaaS tools. If your product cannot be discovered, parsed, and ranked by an agent, you are absent from the choice set. Therefore — and this is the leap — every company on the internet needs to ship an MCP server, the way every company once needed an SEO strategy. The framing being used is "MCP is the new SEO," and the implication is that the shift is comparable in scale and urgency.
A portion of this is correct, and the institutional weight behind it has been building quickly. OpenAI and Stripe shipped an Agentic Commerce Protocol on September 29, 2025. Google announced its Universal Commerce Protocol on January 11, 2026, at the NRF retail conference. An enterprise commerce thesis appeared in Fortune in April 2026, written by Visa's CMO, arguing that AI agents are the new customer segment that companies must learn to sell to. Visa's underlying Business-to-AI report, conducted with Morning Consult, found that seventy-one percent of businesses say they would optimize their products for AI agents. Agentic flows are real, the protocol layer is consolidating, and the long tail of routine commerce — paper towels, flight rebooking, basic compliance forms, low-stakes software trials — will move into delegated agentic surfaces. Companies that serve those flows do need agent-readable surfaces, and the ones that don't will lose share to the ones that do. We are not going to argue otherwise.
But the analogy is doing far more work than the underlying fact supports. SEO became important because Google was the front page of the internet, and the front page was a ranked list shown to a human who then decided. The optimization problem was visibility to the human eye, inside a choice set the human could see. The MCP-or-die argument quietly replaces "be visible to a human" with "be selected by an agent." Those are not minor variations of the same problem. They are different problems with different mechanics and different failure modes. The companies that optimize for the second while believing they are doing the first will find, in three years, that they built infrastructure for the wrong reader.
The category error is hiding in plain sight. SEO operated on the presentation layer. The agent thesis operates on the decision layer. The two are not adjacent stages of the same funnel — they are different problems with different actors, and the leverage of optimizing each is calibrated by what the human is doing at that moment.
Inside SEO, the human is the actor. The system surfaces ten options; the human reads, judges, clicks. Optimization for the presentation layer is optimization for human attention. Inside the agent thesis, the human is meant to be absent. The agent is the actor. The system collapses ten options into one, or zero, and the human is told the answer. Optimization for the agent is optimization for being chosen instead of a human choosing.
This is not the same thing. And the difference matters more than the analogy admits. SEO survived its own ugliness — bad results, paid placements, content farms — because the human could see what was happening and route around it. The agent layer hides what is happening. That changes the trust dynamics in a direction the SEO comparison does not account for.
The presentation layer and the decision layer are different problems
Consider what SEO optimization actually meant in its golden age. It meant being on the first results page, ideally above the fold, ideally in the first three blue links. Behind every dollar spent on SEO was an implicit model of human attention — what would a person see, what would they click, what would they ignore. The medium constrained the optimization. There were ten slots on a page, and the human's eyes did the rest of the work.
The MCP analogy assumes the new equivalent is being selected by an agent in the same way one was selected by a result rank. But the operations are not symmetric. The agent does not present a ranked list to a human and let the human decide. The agent reads the rank, applies its own scoring function, and returns the chosen output. The optimization target collapses from "be visible inside the choice set" to "be the answer."
This sounds like a more efficient version of the same game. It is not. Two consequences follow from the collapse. The first is that the failure mode changes shape. Under SEO, a low-quality page was punished by users scrolling past it; the cost of a bad ranking signal was distributed across the choice set. Under agent selection, the cost of a bad output is concentrated on the single answer returned. The human cannot see the runners-up. They can only judge the one they were given. A wrong answer is the entire experience, not part of it.
The second consequence is that the human's relationship to the system changes. Under SEO, the user trusted the medium because they could see what it was doing. Under agent selection, the user must trust the agent's choice without seeing the choice set. That requires a much higher standard of accuracy from the agent, and a much lower tolerance for error. SEO's tolerance for mediocrity was a feature of its transparency. The agent layer does not get that grace, because it does not show its work.
This is the difference between a product designed for an Architecture of Certainty — where the right answer is obvious to a human inspecting it — and a product designed for agent-legibility, where the answer arrives stripped of the reasoning a human would use to verify it.
Even when the choice set is shown, the medium of showing it is not neutral
The strongest version of the MCP-or-die argument anticipates the previous objection. It says: agents do not hide the choice set. They surface their reasoning. The user can ask to see the runners-up. The transparency is built into the protocol.
This is true at the literal level and misleading at the operational one. The medium of agent disclosure is text, almost always, and text is a dense and serial medium in a way a search engine results page is not. A results page presents ten options as a visual array that the eye scans pre-attentively in a few seconds, pulling signals from layout, typography, snippet density, sponsored markers, domain authority, and the geometry of the page itself. Reading ten text descriptions of the same ten options is a sequential operation. It costs more attention, it takes longer, and it strips out the parallel signals that made the visual array scannable in the first place.
That friction is not incidental. It is the new bottleneck. Faced with a choice set rendered in chat form, most users will not work through it — they will ask the agent to choose. The transparency is preserved in a literal sense; the operational result is the same delegation the agent thesis was supposed to mediate. The runners-up are technically visible and functionally invisible.
A results page was self-disclosing — its medium was its disclosure. A chat reply is not. The medium itself demands work to extract the kind of signal a results page surfaced for free, and across the volume of decisions the agent thesis presumes, most users will not pay that cost. They will delegate, and the protocol's transparency becomes a fig leaf.
The principal-agent problem comes with the architecture
Agents are not neutral. They are software systems with training data, ranking weights, and — in time — monetization arrangements. Anyone who has spent five years watching Amazon's search results decay, or Google's first page accumulate ads, knows what happens when an intermediary that initially served the user gradually starts serving its operator. The pattern has had a name since Cory Doctorow's 2022 essays — enshittification — and it is structural, not anecdotal: the platform first courts users, then courts advertisers, then courts itself. The optimization changes underneath the user, and the user's recourse depends on whether they can see it changing.
In an SEO world, that recourse was concrete. A user could scroll past sponsored results. They could try a different query. They could go directly to a brand they trusted. The visibility of the choice set was a check on the medium. In an agent world, the choice set is hidden by design. The user does not know what was filtered, why, or under whose authority. The check disappears, and the only signal the user receives is the quality of the final answer.
This creates a structural fragility specific to agent commerce. A single bad recommendation, in a category that matters, costs the agent disproportionately. The user does not adjust their search behavior. They distrust the agent. And once they distrust the agent on a high-stakes decision, they go around it. They call a friend. They read a review. They walk into a store. Previous attempts to replace human choice with intermediated selection have hit this ceiling, and the ceiling lowers as the consequence of the decision rises.
The argument's defenders often invoke an essay that has become foundational reading in the AI research community — the 2019 Bitter Lesson, which holds that general methods leveraging compute will win out over methods that bake in human knowledge. They are not wrong about that lesson in narrow technical domains. But they are extending the claim past where it holds. That argument was about how AI systems learn, not about whether humans want their high-stakes choices made for them. The first is an empirical observation about model architectures. The second is a claim about preferences, and it is being asserted, not demonstrated.
The implication for builders is the inverse of the MCP-or-die thesis. The companies most exposed to the agent layer are the ones whose product is interchangeable in the eyes of an agent — commodity goods, routine services, fungible software. The companies least exposed are the ones whose customers have learned to distrust intermediaries, because the intermediary cost is highest where the stakes are highest. Building for the second group is not building backward. It is building where the agent layer cannot reach.
The companies worth building win by being trustworthy to humans, not queryable by agents.
Commerce is not one market, and the high-margin half is not for agents
The MCP claim treats commerce as monolithic. It is not. There is a mechanical layer — price, spec, convenience, replenishment — where the choice is largely commoditized and the human's role in it is mostly friction. Agents will eat this layer. The companies serving it should ship MCP servers, and the ones that don't will lose share. That is the part of the take that is right.
But there is a second layer, and it is where most of the high-margin economy actually lives. It is the layer where humans want to be the one who chose. The wedding venue. The therapist. The lawyer. The school for the child. The co-founder. The doctor for the surgery that cannot be undone. The advisor for the money that took a decade to accumulate. This is not sold through an MCP server. It is not solved by being legible to a ranker. The selection happens inside a relationship — between a human and another human, sometimes mediated by a product, but not replaced by one.
This is the distinction we drew at the model level in the calculator and the colleague, and at the labor level in the scribe economy. Most prestigious knowledge work has a mechanical layer and a judgment layer; agents eat the first; humans rarely delegate the second. The MCP claim is making the same conflation at the commerce level. It is true about the calculator and false about the colleague. The error is treating one observation as if it applied to both, and the consequence is that builders will pour effort into agent-legibility for products that the agent layer was never going to decide.
The companies most worth building in the next decade sit on the judgment side of that line — in markets where information asymmetry costs people the most, and where the right product changes the outcome of the decision, not just the experience of choosing.
The SEO analogy was always going to mislead, because SEO solved a problem of attention and the agent thesis is a claim about decisions. Attention is downstream of medium. Decisions are upstream of trust. The first migrates with the interface. The second does not.
AI is not the internet. It is categorically different — it augments decision quality, not inputs. The builders who internalize this will design for the human in the loop, not against them. The builders who do not will spend the next three years optimizing for a reader who never had the authority to choose them in the first place.
A8C Ventures is an AI-native firm building technology for industries where information asymmetry costs people the most.
