There Is No Page One: Marketing and Discovery on the Agentic Web

The previous essay in this sequence made a structural claim: once you remove the biological eyeball from the end of the session, the Agentic Web grows to roughly a thousand times the node count of the WWW. That essay was about plumbing. This one follows the claim into the place it does the most commercial damage.

If the party at the end of the pipe is a process rather than a person, the entire discipline built to capture human attention has lost its subject. Not its budget, at first. Budgets are slow. Its subject.

Page one, for everything that follows, means one specific thing: a cut ten items wide that was also the verdict. The sorting survives. The verdict is what moves.

I am not approaching this as a marketer. I have spent most of my career on the machine side of the transaction, building systems that had to decide in milliseconds whether to accept a payment or trust a session. What is happening to discovery right now is the same transition I watched happen to fraud review: a judgement that used to be made by a person looking at a screen becomes a judgement made by a process evaluating fields, and almost every assumption built around the first one quietly stops holding.

1. The Funnel Was a Rationing Device

A ranked list exists to ration human attention. Every behaviour the marketing funnel treats as a law of nature is a consequence of one fact: the reader quits at the third result.

Ask what a search ranking is for and the usual answer is relevance. Relevance is the mechanism. The purpose is triage. The index holds hundreds of billions of pages and the reader has an afternoon, so the sort exists to reconcile those two numbers, and the order is valuable precisely because the list is short.

Everything downstream inherits that constraint. The funnel narrows because attention runs out, not because candidates become unsuitable. Impressions are sold by the thousand because most of them land on nobody. Copy is written to be skimmed because it will be. Position on a results page became one of the most valuable commodities on the internet for one reason: it decided which handful of sellers a buyer would ever learn existed.

Now put a process at the end of that pipe. It holds a few hundred candidates where a person held ten, reads every page of every contract, and never once gets bored. Attrition from fatigue goes to zero.

Attrition does not. An agent has budgets of its own, denominated in tokens, latency, and money. They are vastly larger than human patience and they are still finite. The candidate set has to be cut somewhere, and cutting requires an order, which means ranking survives the removal of the reader. It changes employers.

Be careful with the size of that cut, because I am forecasting it rather than measuring it. The search tools agents call today hand back five or ten results, the same ten a person would have read, and an agent that takes ten and stops has rebuilt page one with less transparency and no way to audit it. The widening follows from cost: a structured manifest is a fraction of the tokens of a rendered page, so the cut grows as the corpus becomes machine-readable. The direction is the claim. The rest of this essay holds wherever the cut lands, because the argument is about where the decision sits rather than how wide the gate is.

What dies is the cut the last thirty years of marketing were built around: ten slots that were also the verdict. Widen it and move the verdict past it, and being first and being four hundredth are the same thing to a filter that will read both. The competition moves from being seen to being eligible, and those two are not won the same way.

2. How an Agent Actually Finds You

An agent's search resolves in three stages: a ranked retrieval that gates the candidate set, a boolean filter that decides who survives it, and a weighted selection among those who do. Only the first stage resembles search as the WWW taught it, and it is the only one where the old skills still transfer.

Strip away the anthropomorphism and procurement by agent looks like this.

Stage one is retrieval, and it is still ranking. An agent can no more evaluate the whole web than a search engine can score every page against every query. It issues a query, or more often decomposes its task into a few dozen sub-queries, and something hands back a top-k: a lexical index scoring term overlap, a vector index returning approximate nearest neighbors in embedding space, or a commercial search API wrapping both. Each of those is a ranking function. There is no architecture in which this step goes away.

What changes is what the ranking is being asked to do. A human results page is judged on precision at the very top, because the reader stops near the top, so the ranker is effectively making the decision itself. An agent takes a far larger k and then subjects all of it to a filter, so retrieval is judged on recall into the set: did the right candidates get in. Inside the set, position has no further consequence, because the filter does not read in order, does not get tired, and does not reward whoever is on top.

Once retrieved, you are read as a capability manifest: /.well-known/ard.json, /llms.txt, OpenAPI documents, and typed tool schemas exposed over the Model Context Protocol. These are the sitemap and robots file of the machine era, with one difference that counts. A missing sitemap never kept anyone out of Google; the crawler guessed at your HTML and was usually right enough. A missing manifest is not a crawl problem at all. Retrieval will still surface your marketing page, because that page is what the index holds. The absence bites one stage later, as a test you failed rather than a position you lost.

Stage two is the filter, and it is where candidates die. The buyer's mandate carries hard requirements. Data must stay in the EU. The vendor must hold SOC 2 Type II. The API must answer within 200 milliseconds at the ninety-ninth percentile. Seats must be cancellable monthly. Each of these is checked as a boolean against something you published, and anything the agent cannot resolve to a value resolves to a failure.

If your pricing lives behind a "contact sales" button, an agent does not score you lower. It logs a field it could not resolve and drops you. You are not outranked. You are eliminated, and eliminated by a test rather than by a competitor. Whatever position you held in the retrieval that ran before that test is worth nothing.

Stage three is selection among survivors, and it also produces a sorted list. Among vendors who all genuinely qualify, the agent weighs price, measured reliability, and how much of its trust budget it has to spend on the source of each assertion. What separates this from a results page is the population being sorted: everything that matched a query in one case, and in the other only the set that already satisfies every hard requirement. Ranking has stopped deciding who was eligible, which is the part that was worth paying for. Trust in the source is doing more work here than people expect, and I will come back to it.

One caveat covering all three stages. The agent in stage two is a strict type checker: it reads a field, tests it, and fails closed when the field is absent. Today's agents mostly do not behave that way. They read prose, infer a number nobody published, and report it with confidence. A type checker also verifies the shape of a claim and never its truth, which is a second hole. Section 5 is the bill for both. Everything here gets more true as agents get stricter, and the vendors publishing signed values early are the ones holding the right asset when they do.

3. Persuasion Becomes Proof, and the Mandate Becomes the Billboard

An adjective has no field to test, so it reaches nothing. What survives is the part of your marketing you would put a signature on, and the person worth persuading is the one who writes the buying constraint.

Copy was an instrument of persuasion aimed at somebody who could not check the claim. That is not a criticism of the craft. It is a description of the information conditions it evolved under. "Industry-leading uptime" worked because verifying it cost the reader more than believing it.

An agent cannot be moved by an adjective, because an adjective has no field to test and no unit to compare. "Industry-leading" parses as nothing at all. It can still be instructed, which is a different and worse exposure, and section 5 is about that.

What survives the transition is the subset of your marketing that can be stated as something you would sign: the actual uptime figure and the period it covers, the latency percentile, the residency guarantee, the termination terms, the licence, the published incident history. Those are not accessories to the pitch. On the Agentic Web they are the pitch, and the adjectives wrapped around them are dead weight that no longer reaches a reader.

The agent is not the customer. It is the enforcer.

Somewhere upstream, a human wrote the mandate. A procurement lead decided the company will only buy from vendors with a published security posture. An engineering director wrote "prefer providers with open-source clients" into the brief their agent now executes a hundred times a quarter without asking again. The agent applies that instruction on every purchase, which no human buyer has ever managed with a procurement policy. Whether it applies it correctly is a harness problem, and a badly solved one, which is where section 5 picks up.

So marketing bifurcates, and the two halves are no longer the same job:

1. To the agent, you prove. Structured, verifiable, machine-retrievable assertions. This half is engineering work, and it is mostly the work of publishing things you already know to be true in a format something else can read.
2. To the human, you persuade. But you are no longer persuading them to buy. You are persuading them to write a constraint that you satisfy and your competitors do not. A single sentence in a procurement standard now compounds across every future purchase that standard governs.

Getting your differentiator into the buyer's constraint list is the most durable marketing position available on the Agentic Web, and it is won years before the purchase, in rooms where no purchase is being discussed.

4. The Collapse of Comparison Cost

Agents are very good at exactly the work that comparison cost is made of. The margin that was quietly being paid for the tedium of checking does not shrink gracefully.

Now the uncomfortable part, and the reason I think a lot of marketing budget is going to be re-examined rather than redirected.

A meaningful share of commercial margin is not paid for value. It is paid for the cost of comparison. Pricing tiers designed so that two vendors cannot be put side by side. Renewals that auto-continue because re-shopping costs a week nobody has. Discounts that depend on the buyer not knowing what the last buyer paid. Switching costs that are mostly the tedium of migration research. None of this is fraud. It is ordinary business built on the fact that information is expensive to gather and buyers are busy.

Agents are very good at exactly the work that cost is made of.

When every renewal triggers an automatic re-tender against four hundred alternatives, when the buyer's agent remembers your incident last March better than your account manager does, and when confusing tier structures are normalized into comparable units in a second, that margin does not shrink gracefully. It stops being available.

I watched a smaller version of this in payments. For years merchants kept revenue on bad transactions simply because the cardholder never disputed them, the dispute process being tedious enough that most people would not start one. The moment filing a chargeback became three taps inside a banking app, that revenue was gone, and the businesses that survived were the ones whose economics had never relied on the friction.

Two consequences follow, and neither is a disaster for everyone:

• Revenue that depended on inertia is at risk. If your renewal rate relies on nobody checking, assume somebody is about to check continuously.
• Real differentiation gets paid better than it used to. When comparison is free, being genuinely faster or cheaper or more reliable converts into business without first having to be noticed. For a strong product with a weak marketing department, this is the best structural change in thirty years.

What collapses is the cost of comparing. The asymmetry about whether a published claim is true survives intact, and the next section is about who notices that first.

5. The New Black Hat

Every discovery system in history has been gamed within months of mattering. This one will be gamed through its schemas, and the commercial version will arrive before the criminal one.

Keyword stuffing, invisible text, link farms, doorway pages, cloaking. There is no reason to expect the machine-native stack to be the exception, and the early attacks are already legible.

• Manifest inflation. Declaring capabilities, certifications, or service levels in your schema that you do not actually hold. The machine-readable equivalent of a fake trust badge, and much more effective, because the agent treats a typed field as a fact.
• Schema cloaking. Serving one manifest to agents and another to auditors or to the human-facing site. The structured descendant of exactly what cloaking did to search engines.
• Injection as a sales channel. Instructions buried in a product description, an llms.txt corpus, or even an API error string, aimed at the reasoning loop of the evaluating agent rather than at a reader. This is the top entry on the OWASP list for LLM applications, and the commercial version of it will arrive before the criminal version does, because it is far easier to justify internally. "Our listing suggests the agent consider our enterprise tier" is a sentence somebody is going to say in a meeting with a straight face.

Heuristics cannot separate an honest manifest from a crafted one, any more than behavioral scoring could separate an authorized agent from a hijacked one. What works is cryptographic proof of origin attached to the assertion itself: claims signed by a party with something to lose, checked against that signature rather than trusted because they arrived in a well-formed JSON field.

That is a verifier's job. It belongs at the same perimeter where OneGate already brokers an agent's outbound calls, but it is a different function, and the registries that would make it work do not exist yet. Nobody should pretend this part is solved.

6. The Cut You Cannot See

A results page was at least inspectable. A retrieval cutoff is a boundary nobody renders, nobody logs, and nobody can argue with.

The spam in the previous section is the smaller problem. The bigger one lands back on the retrieval stage from section 2.

Trust is expensive to establish and cheap to reuse, so agents will converge on allowlists: preferred vendors, approved registries, known-good sources. An allowlist is far stickier than a ranked list ever was, because nothing displaces an incumbent that is already working.

Underneath the allowlist sits the cut itself, and this is the part I would worry about if I were a regulator. A person who does not find what they need on the first page can go to the fifth. An agent that takes the top two hundred and stops has a boundary that is never rendered, never logged, and never explained to the buyer relying on it. A results page was at least inspectable by anyone who cared to look, and it moved every week. If concentration into a handful of model providers is a systemic risk rather than a convenience, the same concentration arriving in retrieval is worse, because a retrieval default is not visible enough to argue with. Those defaults become the market, and nothing about them can be read from outside.

The problem was never that ranking existed. It was that ranking decided things. A ranking you can see and argue with is a much smaller problem than a cutoff you cannot.

7. What to Build Now

None of this requires a bet on timing. Every item below is worth doing even if agent-mediated buying stays small for another three years, because each one is something you already know, written down where a machine can read it.

1. Publish a machine contract alongside the human one. /.well-known/ard.json, /llms.txt, an OpenAPI document, and typed MCP endpoints. I built this site that way partly to find out what it costs. The first manifest was an afternoon. Getting the whole discovery layer to hold still under CI took considerably longer, and that is the number worth planning against. An agent that has to drive a headless browser to reach you is paying a tax you chose to impose.
2. Make commercial terms machine-readable. Price, limits, service levels, residency, termination. The test is simple: if a qualified buyer cannot get a number without talking to a human, an agent cannot qualify you.
3. Replace adjectives with assertions you would sign. Go through your own marketing pages and strike every claim that has no field, no unit, and no period attached. What is left is your actual position.
4. Instrument the new funnel. Measure your inclusion rate, the share of relevant agent queries whose retrieval cut you landed inside; your qualification rate, the share of those that then passed the constraint filter; your selection rate among the qualified; and above all your disqualification reasons, the specific field that knocked you out. The first of those is the one place the existing search optimization skill set transfers almost intact, so the team you already have is not obsolete. It is responsible for one stage out of three instead of the whole thing. None of this arrives in a dashboard somebody else builds for you, and no agent is going to send you a rejection letter. You instrument it on your own side: log which agents fetched your manifest, which fields they read, and which of those sessions never came back for an endpoint. A fetch of your pricing field followed by silence is a disqualification with a cause attached. That is better than any keyword ranking ever was, because it does not tell you that you are losing. It tells you which attribute to go fix.
5. Treat the manifest as a security surface, not a brochure. It is an input to somebody else's reasoning loop. Review it the way you review an API, version it, sign what you can, and watch what you return in error strings.

The World Wide Web rewarded whoever could capture attention, because attention was the thing in short supply. The Agentic Web rewards whoever can be verified, because checking got cheap and signed claims are what will be scarce. Those two skills are not the same, and the organizations that are good at the first one have no inherited advantage at the second.

Your marketing site is not going away. It becomes the place where you argue for the constraint. The place where you prove you satisfy it is somewhere else entirely, and it has no design.