Field Notes · 8 min read ·
Second-prompt optimization, part 2: the internet's parallel evolution
Nobody buys from an AI on the first question. “Recommend a hotel in Izmir” puts you on the list; “two nights in September, this is my budget” closes the booking. The first question is solved. The second is not, because the internet is only now being rebuilt for machines.
Your website was built for a pair of eyes. The thing reading it now does not have any, and it is not reading well: across billions of crawler requests, Vercel and MERJ found that 34.82% of ChatGPT's fetches land on a 404. Roughly a third of the doors the assistant tries while looking for you are already bricked up, and nobody sends you a report about it.
That is not a ranking problem and it is not a content problem. The internet is being built a second time, in parallel, for a reader who is not human, and the second build sits on the same rung the first one started from. Read-only. A read-only web could be searched; nothing could be bought on it. Getting from one to the other took another decade and a second layer of protocols, and we are standing in the equivalent decade right now.
Which is why the question being sold to you is the wrong one. Run a boutique hotel in Alaçatı and appearing in the list is prompt 1: it puts you in the running. Being the right answer for somebody with two nights in September, a fixed budget and strong feelings about breakfast is prompt 2, and prompt 2 is where the booking happens. Part 1 said it in one line: nobody has ever bought a sofa at prompt 1.
Prompt 1 is a directory problem. Prompt 2 is a protocol problem, and the protocol does not exist yet.
This post is about why prompt 2 is still unanswerable, and about the single job a read-only era actually rewards.
The internet has done this before
Berners-Lee's 1989 proposal to CERN did not describe a publishing medium. It described a way to manage information: notes joined by links, moved over shared protocols, with the distance between two people taken out of the equation. The problem he was solving was a laboratory that lost what it knew every time somebody's contract ended.
Then somebody had to read it, and the somebody was a person. HTML exists for that and for nothing else, which is to put information in a shape a human eye can scan. That is the read-only web, and it was never a technical ceiling, because the first browser Berners-Lee wrote was also an editor. Writing was there from the beginning. Reading is what spread.
The .com bubble opened the read-write era and the whole thing changed shape: accounts, posts, orders, payment rails, a digital economy nobody has finished counting. Hardware pushed it further. When the phone arrived we did not extend the web, we mirrored it, and every site got a second parallel version living inside an app.
Four eras, one constant. The user was always a person, and every layer of it was tuned and re-tuned for a person, which is why your site has a navigation bar, your catalogue has photographs, and your pricing page has a table somebody spent a week aligning. So how efficient is any of that for a machine?
None of it. Not one layer.
AI's first replacement was not your job. It was your interface.
Thirty or forty years went into a body of work whose only purpose was letting a person talk to a computer: markup languages, forms, buttons, menus, interface guidelines, app stores. Set against the history of writing things down, that stretch is shorter than a single human life, and it was always going to be forgotten if it ever finished the job and produced a successor.
It did. The successor is AI, and the most useful thing about it is not any one capability. It is that it translates: our way of thinking and our way of saying things, run through mathematics and probability, carried down from words to code and from code to ones and zeros.
Before it comes for anybody's job, it replaced the decades of work we did to make computers usable.
How young that body of work really is, and why we were never meant to be doing it, is a separate post: the oldest tool in a white-collar day is forty-seven years old.
If nobody has to learn the button, the button does not have to exist.
AI's internet is at read-only too
We are now redesigning the internet, and operating systems with it, so that an AI can use them. Stage one is read-only again: an information layer, and the problem of reaching correct information through it. None of that ordering is a surprise, because you cannot let something act on the world before it can read the world correctly.
Put the two evolutions on top of each other and the rung we are standing on is obvious:
That rung is not finished, so the machine is currently feeling its way around a house built for human eyes. Vercel and MERJ measured this across billions of crawler requests and found that no major AI crawler runs JavaScript. GPTBot and ClaudeBot download your JavaScript files without ever executing them.
Your best page can render perfectly in a browser and come back empty from the server, and nothing warns you that it happened, because every test a human runs on it passes. A page like that is not a page. It is a rumour.
Hallucination is not a character flaw, it is an unfinished read layer
Four years in, the loudest complaint about these models is still hallucination. Get a concert date wrong, or a company's phone number, and every impressive thing the model can do stops counting for the person who asked, because they now have to go and check the rest of the answers too.
The cause is not mysterious: researchers at OpenAI and Georgia Tech argued in 2025 that models are rewarded for guessing when they are unsure, because training and evaluation penalise "I don't know" and pay out for an answer that looks right. The model behaves like a student who would rather guess than leave the question blank. What is missing is not character, it is a source: when a model cannot find a readable, correct record about you, it does not go quiet, it fills the gap.
When the assistant invents your phone number, the model is not the one who loses.
The industry measures visibility because visibility is what it can measure
AEO and GEO have covered real ground in a very short time, and the academic starting gun is recent. The paper presented at KDD 2024 calls itself the first framework for optimising content visibility inside generative engines, and measures its own methods lifting visibility by up to 40%.
Since then you can find out whether the engines mention you, which is worth having. What gets pushed, how hard, and how far a brand can climb out from under the old giants still moves with a consensus that changes every month. New integrations keep landing, which both strengthens the field and forces it to stay quick. The maps integration we wrote about in part 1 was one of them, and it flattened an entire category in a single move.
Prompt 1 gets asked. Prompt 2 gets decided.
Read these two out loud, one after the other. First:
I'm going on holiday this summer. Can you suggest sites where I can compare the options?
Visibility matters enormously here. Classic search optimization, how fresh your site and your content are, and what the public says about you all feed into it, each one somehow more decisive than the last. The assistant does what you expect and hands over a flat list:
List[ … ]
But is that where the buying decision gets made? Second:
Right. Now split those by city, factor in what people said in the reviews, and you know me, so pick me the best value for money.
Those are not the same question at two volumes. The first is about the world, and the second is about the person asking, which is exactly why its answer is not sitting in any list. The NBER study that tracked ChatGPT use from November 2022 to July 2025 found that the three commonest uses, practical guidance, seeking information and writing, account for close to 80% of conversations, and that non-work messages climbed from 53% to more than 70%. People use this thing as an advisor who knows them, not as a search box.
An advisor gets a follow-up, and the follow-up is the one you cannot answer. The hotel above can sit in every list an engine produces and still have nowhere to put the sentence that closes the booking, because that sentence is not a fact about the hotel. It is a fact about the guest, and it does not exist until the guest has said it out loud.
In a read-only era there is exactly one job, and it is being readable
Being the answer to the second question is the whole skill.
That means your brand has to be visible to an AI about its own business, fully and objectively, the way a read-only MCP is. Anthropic published the Model Context Protocol in November 2024 as a universal, open standard for connecting AI systems to data sources, built to kill the problem where every new source needed its own custom integration. Its competitors settled the question of whether this was one vendor's enthusiasm: OpenAI adopted it, and Google DeepMind then said it would add MCP support to Gemini and its SDK.
The second job is getting the old-generation site you built for people out from under the JavaScript barrier and open to the machine as well. Vercel's advice from the same measurement runs to one sentence: server-render anything that matters. It changes nothing about how the page looks to a customer, and it decides whether the page exists at all for the reader you cannot see.
The third job is the one everybody skips, and it is being transparent and objective. In a read-only era what your brand says is what the model learns, so a record you inflated becomes the source of the hallucination you were trying to fix. That is what we are building at EOMA.
Everyone is optimising the last page of the first evolution. The ticket is sold on the first page of the second. Not first-prompt optimization. Second-prompt optimization.
Read-write is next
So what happens when the AI internet reaches its read-write era?
The next post gets into it properly, but here is the spoiler:
Mate, I'm ordering food. I have four apps and I cannot keep the offers straight. Can you compare the deals and the restaurants across x, y, z and w, and give me a shortlist?
In that sentence the assistant has stopped reading and started acting for you. The infrastructure was not built on that assumption. Google's Agent Payments Protocol, announced in September 2025 with more than sixty organisations behind it, opens by saying that today's payment systems assume a human is clicking "buy" on a trusted surface, and that autonomous agents break that assumption. The read-write era is being wired right now, protocol by protocol, the same way the read-only one is.
Which is where part 1's promise lands. Part 3 is the third prompt, the one that comes after they have chosen you, in a world where the assistant is no longer recommending on your behalf but sitting down at the table on the customer's.
In the first evolution somebody else decided how your brand would be read. The second one is still on its first rung, and that decision is still yours.