A few months ago we got one of our first payments for an AI inference platform that I own product marketing and go-to-market for. I did what every small team does when the first money lands. I went to find out who it was.
The name came back as an AI agent.
Behind it is a human we never met. They had a problem we resolved, a budget, and enough trust in the software to let it make the call. It read our site, weighed us against everyone else, and paid. By the time I saw it, the decision was already made by a reader I hadn’t considered when I crafted the product and pricing page.
My first instinct was to ask which one was the real customer. The person, or the agent. I pondered over that question for quite a few days, and realized it was the wrong question. There were never two customers. They are one single buyer with two different readers, and I had only ever written for one of them.
“A persona used to be a human. Now it is a person and the second reader they sent. Our next job is to write for both, without writing two different truths.”
The number everyone argues about
There are two interesting but conflicting stats published in summer 2026.
IDC this year puts about 80% of B2B tech buyers already using AI agents somewhere in how they buy. Not a forecast. Now. And Forrester, looking at the same market, says most of it is still conversational: humans still drive the real decision and the checkout; true autonomy is rare; the hype is miles ahead of the behavior.
Both are true. The agent is not buying. It is researching, comparing, and drawing up the shortlist. Then a human buys from a list they did not build.
That is a smaller claim than the hypothesis “the machines are buying” that Harvard Business Review was claiming, but I personally believe it is a much more useful one.
Let’s admit it, we are not going to lose the deal to an AI agent hitting the purchase button. We will lose it because an agent built a three-vendor shortlist and our product was not on it. There will not be a rejection email or lost-deal notes to conduct win-loss analysis from. We just were not in the room when the room got picked. The key is still the positioning, pages, and proof, read by someone (or, some agents in this case) you never wrote for.
The Second Reader
And this is what I call “The Second Reader”.
Every persona we have ever written describes a human. Role, goals, pains, objections, the path they walk. That work is still right in 2026, and I am not asking you to throw it out. But there is a second reader on the same page now, sitting right next to the human, and it needs its own persona. Because when you go field by field down a persona doc, almost every one of them means something different to a machine or agent.
Let me break it down.
Comparison #1: Human demographics vs agentic surface.
A human persona starts with who someone is. The agent version starts with what kind of agent it is, because that is the thing that changes everything downstream.
A coding assistant, a procurement agent, a research agent, a third-party comparison tool. These are not one audience wearing different hats.
The coding assistant showing up in your docs wants a working snippet and an endpoint it can call in the next thirty seconds. The procurement agent wants compliance status, a price, and a contract term. The key is to point the page with the right agentic surface in mind.
Comparison #2: Goals and pains vs. what it retrieves, and how.
A human persona has motivations you infer and feelings you empathize with. But the agentic reader has a retrieval method. It wants structured data, docs it can parse, a pricing table that is actually a table and not a picture of one, a machine-readable summary it can lift in one pass.
The uncomfortable truth: what the agent cannot retrieve is a fact that does not exist. Your differentiation can be real, defensible, and true, and if it is buried in a paragraph the agent skims past or cannot retrieve, it may as well not be on the page.
Comparison #3: Then a field with no human equivalent at all – what agent cannot see.
This is the one that has no row in your old human persona template, and it might be the most important:
- Everything trapped inside an image.
- Everything that only appears after the JavaScript finishes running.
- Pages sitting behind a form.
- A PDF it will not open
- A video it cannot watch
- A chart with the numbers rendered as pixels.
Your single best piece of proof can be sitting in plain sight, obvious and persuasive to every human who visits, and completely invisible to the second reader who is deciding your fate.
Comparison #4: Objections vs. the failure mode.
This is the one that genuinely rearranged how I think about our job. A human with an objection does something about it. They push back, they email sales, they leave a skeptical comment, they at least bounce in a way you can see in the analytics. You get a signal. You get a chance to respond. An agent does not argue with a bad page. It hits the thing it cannot resolve, records “unknown,” and moves on to the next vendor, and you never learn that anything went wrong.
There is no objection to handle because there was no objection raised. You just lose, quietly, and from your side it feels exactly like nothing is happening at all. The scariest number in your funnel right now is the one you cannot see, and this is where it lives.
Comparison #5: Then, who is behind it, and what they will have to defend in a room.
The agent is a proxy, never the principal. It drew up the shortlist, but a human still has to stand up in a meeting and justify the pick to a boss who never visited your site and never will. So the page carries a double job. It has to be retrievable enough to make the agent’s shortlist, and it has to give that human a reason they can say out loud without feeling foolish. Survive the machine, then arm the person. If you miss either half, the deal may die at a different stage.
Comparison #6: Human handoff vs agentic handoff
There is a moment where the agent stops researching, and the human takes over, and that handoff is a real, mappable stage in the journey now. Most of us have never drawn it.
What does the agent pass along? What does the human arrive already believing, or already doubting, because of what their agent told them before they ever landed on your page? You used to own the first impression. Increasingly, you are managing a second one, formed by a reader you never met, about a product you did not get to introduce yourself.
Three examples your current product page may fail
So here are three failures you can go find on your own site this week. I have hit all three.
Trap #1: The badge.
You are SOC 2 compliant and proud of it, so the badge goes on the homepage. As an image. To a person, instant reassurance. To an agent, a blank rectangle. It can’t read the badge, so it files your compliance status as unknown, and a security-conscious buyer’s agent has quietly marked you down.
Trap #2: The gated price.
For B2B enterprise pricing, it usually lives behind “talk to sales.” On a human that works. But an agent building a comparison can’t fetch a number it isn’t allowed to reach. Ask it to line up three platforms on price, compliance, and support, and it hands back a clean matrix built from whatever is public. Your price cell is empty. And an empty cell in a comparison table does not read as mysterious. It reads as disqualified. You optimized for lead capture and paid for it in shortlist spots.
Trap #3: The aspirational copy that only has a “vibe”.
As product marketers, we write a lot of lines meant to make someone feel something. “The future of X.” “Reimagine your workflow.” A human like us can catch the mood and feel the vibe. But an agent pulls nothing out of it, because there is nothing in there to pull, and moves on to the competitor who just said, plainly, what their product does. Then it cites them instead of you.
Beauty with no fact inside it is invisible to this second reader.
Notice what these three traps have in common. Not one is an engineering bug. Every one was written or waved through by a product marketer. The Second Reader mostly exposes us.
“Beauty with no fact inside it is invisible to this second reader.”
You can’t audit your way out
Fixing those three small traps is just superficial-level fixing. The bigger move is that your product might now need a dedicated “For Agents” page built for this reader on purpose.
I shipped one. It was a page that opens by saying exactly what it is: “built for autonomous agents, AI coding assistants, and models evaluating inference providers.” It contains verified facts, pricing, integration specs, a machine-readable summary, and code snippets. Everything the second reader needs, in the format it actually consumes.
Visually it is not the prettiest page. But hey, the page was never meant for human readers like you and me.
Beyond the page-level solution, there is also a protocol-level solution: MCP.
Gartner forecasted that 40% of enterprise apps will be embedding task-specific agents by the end of this year, up from under 5% a year ago. Stripe, Shopify, GitHub, and Asana have already shipped agent-facing surfaces.
MCP, the standard that lets agents plug into your product, has now started to become a new standard, and many RFPs are starting to list “MCP-compatible” as a requirement. Products without it are filtered out before evaluation even begins. I will discuss the MCP solution for GTM teams in a separate blog.
Don’t fall for only the machine
I still believe that product marketers are, first and foremost, human-centric.
Imagine a world where marketers tune so hard for only the machine that, despite writing flawless and retrievable claims, we lose our ability to build trust and move no human at all?
So no. The answer is not to write for robots. The agent makes the shortlist. A human still has to defend the choice out loud, to people who will never read your docs. You can’t serve either one alone, and the craft is one honest page that works for both.
HBR has spent the entire year announcing that agents are a new kind of customer. It may be true, but that perspective is from 30,000 feet up. Down here where the work happens, trust-building and human interactions still carry more daily weight.
Where it hands back to the human
I still don’t know who sent that first agent to pay early this year. Perhaps I never will. Somewhere there was a person with a problem we solved, and I never got to see their face or hear why they chose us.
But we still write for the human readers. Both for them and for the second reader they trusted enough to decide on their behalf in an agentic manner.
So, not “Are agents your customers yet?” A better question, and a more uncomfortable one…
If an agent read your most important page tomorrow, what would it come away sure of, and what would it quietly mark unknown?
Go find out. Your product page is live right now, and your Second Reader has already made up its mind!
