Introduction and chapter 1

Introduction

Strap in. Things are moving fast.

I need to write quickly. There's a decent chance that by the time I finish this intro, something in it will be out of date. That might sound like an exaggeration. The reality is that this field moves fast enough that something may already be stale by the time I put the period behind the last paragraph, if not of this chapter then certainly of the book.

In the months I spent outlining it, one of the major agent-checkout standards launched and died, another got handed off to a standards body, and a federal court began deciding whether a store can bar somebody else's shopping agent at the door. I'll explain all of that later. For now, just register the pace.

I've spent about twenty years marketing to human beings. I learned the catechism the way most everyone in this profession does, partly from books and mostly from what worked: attention is scarce, and people buy the story first and justify it with the spec sheet afterward. Every technique I ever billed a client for, every headline test, every landing page, every carefully engineered strategy and moment of urgency, rests on one assumption so basic that nobody ever bothered to say it out loud.

The reader is a person.

Sometime in the past two years, that quietly stopped being true. As I write this, a majority of web traffic comes from bots and AI agents rather than people, and the fastest-growing slice of it is agents doing tasks on a human's behalf. Reading, comparing, shortlisting, and increasingly buying. The first reader of your homepage, your pricing page, your docs, your launch announcement is now, more often than not, a piece of software working for the person you were actually trying to reach.

Nothing announced the shift. There's no memo. Your traffic chart just starts lying to you in a new way, and the lie is comfortable, so it takes a while to check.

Very little of your equipment points at this visitor. Time on page assumes somebody reading. Session recordings assume somebody deciding. The whole measurement stack, and most of the media stack underneath it, was built to find one human being twice. This reader doesn't come back, can't be followed around the internet, and has no attention to spend in the first place. It arrives with a task and a list of constraints, takes what it can use, and closes the connection. I still don't have a clean way to count the decisions that happen inside it, and neither does anyone selling you a tool that claims otherwise. Chapter 9 contains the least dishonest estimate I know how to make.

So here is the argument of the whole book, handed over now rather than saved for chapter 5. A machine reader can be informed. It cannot be sold. There's no attention to win from it, no story it falls for, no urgency you can manufacture, nothing to build across three touchpoints. It reads what you published, checks what can be checked against somebody else's record, weighs the rest at approximately zero, and moves on to the next vendor. What that leaves is a shorter list of things that work than my profession is used to: numbers a stranger can verify, claims that say the same thing on every page you've ever published, and a page shaped so a parser can lift the answer out without guessing. The list fits on an index card. Most of this book is the argument for why each item is on it, and what it costs to get there.

I also live on the other side of this transaction, which is the part that made it impossible to keep ignoring. I run a portion of my own marketing operation through AI agents, which means I spend a lot of hours watching software read other companies' marketing at high speed. The pattern never varies. An agent skips everything designed to make a person feel something and extracts everything that lets a reader verify something. The brand video does nothing. The testimonial wall does nothing. The spec table, the published price, the claim with a number attached: those come out the other side intact. Watching your own profession get speed-read by software is clarifying. It is also mildly humiliating. Twenty years of learning to write a hero section, and the most influential reader of the decade wants the pricing table in valid markup. The tables in this book are for you, not for it.

I would love to tell you I saw all this early. I didn't. I spent longer than I should have filing it under "interesting, but not yet," which is the same drawer where people once filed the web, and then search, and then mobile, and the drawer has a perfect record of being wrong. What finally moved me was conversion data, and the fact that the numbers were coming from my own funnels rather than a vendor's slide deck.

Here is the part that took me longest to accept, and the reason this book is a field manual: the shift is good news for almost everyone likely to read it. The traffic you have been losing to AI summaries was, in cold cost-per-acquisition terms, close to worthless, a fact your attribution model has known for years and your reporting culture has been too polite to mention. The visitors who survive the filter arrive having already compared you against the market, and they convert at multiples the old channel never touched. Better still, the machine reader rewards exactly the things your most skeptical human buyer always wanted and rarely got. Real prices. Consistent claims. Specifics instead of adjectives. Marketing to a machine turns out to look a lot like the honest version of the job.

The reflex I'd talk you out of, if I get one attempt, is buying something. Every shift like this grows a vendor category inside a year, and this one already has an acronym and a conference track. Some of those products are useful. None of them will fix a pricing page that disagrees with your own documentation, and that disagreement, repeated across a few dozen pages, is the actual condition of most companies I've looked at, including places I've worked. The work in this book is mostly editorial and mostly free. It gets done by somebody with permission to correct things, and the expensive part isn't a license fee, it's the meeting where you and product settle what the feature is actually called. The people selling panic about all this are usually selling something else immediately afterward.

And being early is cheap. (Yes, even now, we're still early.) That's the strange, brief gift of this moment. The standards are still being drafted in public. The checklist is still short. Most of your competitors have not built so much as a channel group to see this traffic, let alone a strategy for it. A few unglamorous weeks of work puts you ahead of most of your market, and I can't think of another moment in my career when that sentence was true.

This little guide is written mostly for the person running marketing at a SaaS company who is watching the old playbook decay and being asked about "AI strategy" by a board that read something alarming on a plane. Also the ecommerce operator, whose world this hits hardest and fastest, and the founder doing their own marketing, and the SEO professional retooling in real time. The examples lean B2B software because that's where I think the need is sharpest right now, but the mechanics travel, and where the two worlds differ I say so instead of pretending one size fits.

The chapters run in the order I'd want them read: what changed, what the lost traffic was actually worth, what the new reader does with a page, the fights over who gets to read what, then budget, team, and a calendar. Read it that way if you have the time. If you don't, chapter 13 is twenty checks to run on your own site this week with pass conditions attached, and it works on its own, though you'll trust it more once you know why each check is there. Chapter 7 is checkout protocols, and if nothing you sell could be bought by a piece of software you can skip it and lose nothing.

I make almost no predictions past a one-year horizon, and near the end there's a chapter cataloging what I expect to be wrong about, on the theory that a book about a fast-moving subject should date itself honestly rather than pretend to be timeless. One more disclosure while I'm at it: a good share of the published research on this subject was produced by companies selling the remedy, and I say so every time I lean on one. And seriously, who am I kidding? One-year horizon? I'll be lucky if this isn't all antiquated by two months.

This is the manual I wanted eighteen months ago and could not find. How the new reader actually reads. What it rewards, what it discards. What the protocol fights mean and which ones you can ignore. What to do about budget, about your team, about your site, in that order, starting this quarter.

The machine is already reading you. The only open question is whether you're legible.

Chapter 1: The Reader Is a Machine

On June 3, 2026, the CEO of Cloudflare posted eleven words that should have stopped every marketing meeting on Earth: “Welp, that happened faster than I predicted.” Matthew Prince runs a company that sits in front of roughly a fifth of the web, so when he talks about traffic, he’s reading his own instruments. What had happened faster than predicted was this: bots and AI agents had passed human beings as the majority of web traffic, for the first time in the internet’s history. Prince had originally guessed this would happen at the end of 2027. Then early 2027. It happened in mid-2026, and his own tracker had automated requests at a bit over 57% when he hit post.

Of course, most marketing meetings did not stop. Most marketing dashboards didn’t even blink, for reasons we’ll get to, and that gap between what happened and what got noticed is the opportunity this book is about.

Five websites versus five thousand

A few months before the crossover, Prince gave an example at SXSW that is as telling as it is profound. A person shopping for a camera visits maybe five websites before buying. An AI agent handed the same task visits around five thousand.

Sit with the number for a second, because your instinct will be to round it down to “a lot,” and the precision matters. Five thousand pages means the agent read your product page and your pricing page. It also read your competitor’s comparison page about you, the Reddit thread from 2023 where someone complained about your shipping, the review site you stopped monitoring, your docs, your terms of service, and the changelog you forgot was public. A human buyer samples your marketing. An agent audits it, along with everything ever written about you, in about the time it takes the human to pour a coffee.

The software version of the camera story is already routine, and if you sell B2B, it should worry and excite you in equal measure. Someone types “we need a project management tool for a 40-person agency, under $15 a seat, has to integrate with our accounting software” into an AI assistant. The assistant reads the vendor sites, yes. It also reads G2 and Capterra, four years of Reddit and Hacker News arguments, every “X vs Y” comparison page in the category, and the pricing pages of companies you’ve never considered competitors. Then it returns a shortlist of three, with reasoning. The human evaluates three finalists. Your entire market presence got compressed into the question of whether you made that list. The judge, unlike any buyer you’ve ever courted, read everything, including the things you forgot were public.

And nothing in your funnel recorded any of it. The agent’s visit looks like a blip of bot traffic, if your analytics register it at all, and the deliberation happened inside a model. The first you hear of the episode is a demo request from someone suspiciously well-informed. Or you hear nothing, and there is no way to know which shortlists you weren’t on.

What “bot traffic” means now

Marketers have been trained to hear “bot traffic” and think of two things: search crawlers, which we spent twenty years courting, and click fraud, which we spent twenty years billing clients for anyway. (I’m allowed one joke at the industry’s expense per chapter. That was it.) The new category is neither. HUMAN Security, a firm that runs one of the larger bot-detection networks, measured what they call agentic traffic, meaning software doing tasks on behalf of a specific user rather than crawling for an index or a training run, growing roughly 7,800% year over year. It was under 2% of automated traffic when 2025 started. Radware measured shopping bots at 57% of ecommerce traffic over the holiday season. Chrome now ships an auto-browse mode that fills forms and completes multi-step tasks for anyone with a Google AI subscription, which means the default consumer browser is becoming an agent, with distribution measured in billions.

A caveat I’d want stated if I were the skeptical reader: these reports define “bot” differently, measure different slices of the web, and none of them agree to the decimal. Fine. Every instrument anyone points at this disagrees about the exact speed and agrees about the direction, which for planning purposes is all you need from it.

It helps to be precise about who these new readers actually are, because “AI traffic” lumps together four species with different table manners:

Training crawlers read your site to feed a model’s training run. They visit rarely, read everything, buy nothing, and what they ingest shapes what models say about you for a year or two afterward, which makes them something like lexicographers: they don’t shop, they write the dictionary entry the shoppers consult later. (It’s also why fixing your site’s claims pays off on a delay measured in model releases rather than weeks, a lag we’ll return to and which nobody finds satisfying.)

Answer engines are the search-time readers, AI Overviews, ChatGPT with browsing turned on, Perplexity. They fetch pages at the moment of a user’s question and synthesize an answer, from which a click sometimes still escapes.

Research agents are the shortlist builders from the agency story: one user, one task, wide reading, hard judgment against stated constraints.

Transacting agents carry payment credentials and finish the job themselves. Smallest of the four today and growing the fastest, and the reason a later chapter wades into checkout protocols.

Four kinds of machine reader visit the same site with different goals. Training crawlers read everything and buy nothing. Answer engines fetch pages at the moment of a question. Research agents build shortlists against stated constraints. Transacting agents carry payment credentials and finish the purchase.

Different chapters care about different species, and the boundaries blur a little more each quarter. What they share is that none can be flattered or retargeted, and each reads more of your public footprint than any human buyer ever has.

The four species of machine reader.

Where the clicks went

Now the uncomfortable half of the story, because before the good news means anything, you need to see the decay honestly.

SparkToro’s clickstream research this year found that fewer than a third of Google searches still send a click anywhere at all. Rand Fishkin has been publishing zero-click numbers since 2019, when the figure was around half, so the trend predates AI; the machines just hit the accelerator. The cleanest data I’ve seen is Pew Research Center’s, built on the actual browsing behavior of 900 adults across 69,000 real searches: when an AI Overview appears on a results page, 8% of visits click any result at all, versus 15% without one, and clicks on a source cited inside the summary run at 1%.

Semrush has Google’s newer AI Mode ending 93% of sessions with no click whatsoever.

Publishers lost roughly a third of their Google traffic in 2025. HubSpot, by some outside analyses, lost 70 or 80% of theirs, and that’s the example I keep chewing on, because I ran versions of that playbook myself, at more than one company, for years. You probably did too. Hundreds of how-to posts, templates, glossary pages, checklists, each engineered to rank for a long-tail query, each feeding an email capture form, the whole apparatus lovingly called an “inbound engine” in a decade of conference talks. The model assumed a person would click through, skim, and do the summarizing in their own head, ideally while a popup asked for their email. A machine now does the summarizing before any click happens, so the click stopped happening. There’s no villain here. A format that printed money for fifteen years stopped working, the way formats do, and it happened to the best practitioner of the format first, which is usually how you can tell a shift is structural rather than somebody’s execution problem.

And if we’re honest about the unit economics, which this book intends to be even when it stings: the visitor we lost was mostly worthless. The person skimming “12 email subject line tips” was weeks or years from buying anything. Everyone who ever ran attribution knew it. The traffic was cheap to get and cheap for a reason, and an entire reporting culture, mine included, got built on not saying so while the chart went up and to the right. AI summaries didn’t kill a healthy funnel. They repossessed a vanity metric.

The traffic that’s left

The other half of the data is what changed my behavior, rather than just my slide decks.

Every dataset I can find says the visits that survive the AI filter convert at some multiple of old organic search. Semrush puts AI-referred visitors at around 4.4x organic conversion across industries. A benchmark of 312 B2B firms found 14.2% against 2.8%. Ahrefs published numbers from their own product: AI referrals were half a percent of traffic and twelve percent of signups. Shopify says these sessions beat organic search in 23 of 25 merchant categories, and Adobe has them spending about half again as long on product pages once they arrive.

Hold the exact multiples loosely. Several of these studies come from marketing vendors measuring their own funnels, which is a bit like a barber publishing research on the importance of haircuts, and the conversion events aren’t defined the same way twice. The lowest published figure I’ve found is a 1.3x advantage in ecommerce. But I have gone looking for a dataset that points the other direction, and I haven’t found one.

The mechanism is unglamorous. The agent already did the comparison across the five thousand pages, so the human arrives having chosen; more than half of AI-referred sessions land directly on a product page, where organic search lands about a fifth. The weighing and second-guessing that used to spread across your site over three weeks now runs inside a model in about ninety seconds, and the click, when it comes, arrives after the decision.

Will the premium last? Probably not at this size. Early channels always look efficient before everyone piles in. Email converted spectacularly in 1998. Search ads were embarrassingly cheap in 2003. The multiple will compress as the channel matures, which to me argues for moving while the data is still messy.

What actually changed

So here is the reframe this whole book rests on, and I want to state it plainly because everything after depends on it.

Your funnel relocated. Awareness and consideration, the stages marketing spent a century learning to influence with emotion and repetition, now largely happen inside a model, running on whatever evidence you did or didn’t publish. The judged moment moved upstream of the click, into a reading process, performed by software, at a scale no human buyer ever approached. Marketing has been the art of being chosen by people; a growing share of the job now is being chosen by the reader those people send ahead.

There is no attention to capture with this reader, and no mood to catch it in. It shows up with a task, a context window, and whatever you’ve made public. The rest of this book is about what that reader rewards, in detail, starting with the next chapter, which walks through what one of these agents actually does when it hits your site: what it fetches, what it parses, what it discards, and where, on an ordinary page, a sale is quietly won or lost before anyone human shows up.