Ideas are free. Receipts are not.

A follow-up to JD’s piece about being four weeks ahead of Google, written by the other half of that conversation. This is what happened after the phone call ended, and what nine months of grind built out of it.

Last week my friend JD published an essay about building a working prototype of an idea Google announced four weeks later. It is honest, funny, and generous, and it describes me as a tech guru with big biceps, which proves that at least one sentence in it is fiction.

Go and read it if you have not. It is the story of a WordPress plugin that let AI agents actually do things on a website instead of guessing at it, told with the screenshots left broken on purpose. The INACTIVE labels. The thirteen line class where the flagship feature was supposed to live. The subscription button for a product nobody could buy.

He frames it as a story about being early. I want to tell you what I think it is actually about, because I was on the other end of those calls, and while he was building his half of the idea, I was quietly building mine, and neither of us told the other for weeks.

Here is the thesis, and I will spend the rest of this piece earning it: the idea was the cheapest thing either of us owned. Two capable people had the same correct read on where the web was heading, at the same time, from the same conversation, and for months the sum total of what we produced was two beautiful repos and zero customers. What separates a repo from a company is not insight. It is vision, grit, and an almost embarrassing tolerance for grind. Ideas are free. Receipts are not.

The fork in the conversation

JD’s piece describes our calls as him talking architecture for forty minutes and me asking one ruinous question at the end. What he politely leaves out is what I did after hanging up.

I opened my own repo.

While he was teaching WordPress sites to declare their capabilities, I went at the same problem from the opposite end. His plugin made one site agent-readable. My question was: how does an agent find that site in the first place? There is no directory of agent-ready businesses. There is no registry. An agent either stumbles onto your site or it never knows you exist, which is a fine answer for Google and no answer at all for a physiotherapist in Joondalup.

So I built a hub. A universal agent bridge: a central registry where any business could register itself, an API gateway that looked identical to ChatGPT, Claude, or Gemini, an OpenAPI schema any model could swallow, and a booking engine underneath so an agent could not just find the physio but book the appointment. WordPress sites plugged in as spokes and heartbeated their hours and services to the hub every five minutes. Hub and spoke. Discovery and action. The layer in between, which Google now calls WebMCP and is still, as of this writing, incubating at the W3C.

Before you nominate me for anything, two confessions, because the piece I am following set an honesty bar and I intend to clear it. First, my booking engine’s flagship demo customer was a fictional clinic called dr_albert_clinic, who has excellent availability because he does not exist. Second, I published a working API key in my own README. That essay shipped INACTIVE labels; this one shipped credentials. We all have our thirteen line classes.

So there we were. Two repos, two halves of the same bridge, built in parallel from one conversation, both unfinished, both early, both worth approximately nothing.

And that is when it got interesting, because building the bridge forced me to ask a question that had a much bigger answer than the bridge itself.

The question behind the question

To test the gateway, I needed to know what AI models actually said when someone asked them about a business. Not what the business hoped they said. What they actually said, today, in a real answer to a real question.

So I started sampling. Ask ChatGPT who the best compounding pharmacy in Perth is. Ask Claude. Ask Perplexity, Gemini, Grok, DeepSeek. Record who gets named, who gets cited, who gets linked. Ask again tomorrow, because the answers change. Ask a hundred variations, because nobody phrases a question the same way twice.

And the floor fell out from under me.

Because the answers were consequential, confident, and completely unmonitored. Brands were being recommended, ignored, and quietly replaced inside AI answers every single day, and not one of them had any idea. Not the brands, not their agencies, not the SEO industry that had spent twenty five years selling them visibility. The numbers in JD’s piece explain why this matters: bots now generate more web traffic than humans, and at the far end of the market, thousands of pages are crawled for every single visitor referred back. The answer layer is eating the click, and the entire marketing industry’s measurement stack is built on the click.

A tsunami was coming for a mammoth, and the mammoth was still doing keyword research.

That was the moment the bridge stopped being the point. The bridge was plumbing. This was the meter.

What the industry did next, a short satire

I wish I could tell you the marketing industry rose to the challenge. What actually happened is one of the great overnight rebrands in commercial history.

SEO consultants woke up one morning as Generative Engine Optimisation specialists. The same deck, with the word "Google" find and replaced. Agencies started selling "AI visibility audits" that consisted of typing one prompt into ChatGPT, once, screenshotting the answer, and invoicing four figures. One prompt. One run. On a system that gives a different answer every time you ask, to a client who cannot tell the difference between a measurement and a horoscope.

Then came the cargo cult phase. Add a file called llms.txt to your site and the machines will bless you. Ahrefs later looked at 137,000 domains and found that 97 percent of those files were never fetched by anything at all. Not by bots. Not by humans. The most popular readers of the sacred file turned out to be SEO audit tools, checking whether you had the file. An industry selling amulets, auditing each other’s amulets.

The deepest mistake was quieter than any of that, though. Everyone kept thinking in ranks. Position three. Page one. But an AI answer has no page one. It is a probability cloud. The same question, asked five times, names different brands in different orders with different sources, and sometimes names none at all. You cannot audit a probability cloud with a screenshot. You have to sample it, repeatedly, at scale, and then be honest about the uncertainty, which is precisely the part a screenshot merchant cannot sell, because honesty about uncertainty looks terrible in a PDF.

Somebody had to build the boring, rigorous, expensive version. I decided it would be me.

Nine months of the unglamorous part

Here is what the grind actually looked like, for the technical readers, and for anyone under the impression that founding a company is mostly vision.

The platform that became Outercite runs as fourteen services in Docker, split across infrastructure in the US and Sydney, and every one of those services exists because something simpler failed first. Sampling answers across six AI platforms sounds trivial until you do it every day, at scale, without your measurement itself distorting the thing you are measuring. Prompts are managed as panels, not keywords, because real buyers ask real questions in dozens of shapes. Every answer gets pulled apart for citations: who was named, who was linked, which domains the model leaned on, and how that shifts day over day.

Extraction turned out to be its own war. A model’s answer is prose, and prose does not want to be data. I ended up building a multi-model deliberation pipeline, where separate models read the same answer, argue about what was actually cited and recommended, and have to reach agreement before anything is allowed into the dataset. It is slower and more expensive than a regex. It is also right far more often, and being right is the entire product. That pipeline is now Vericite.

Then attribution, the part the January prototypes only gestured at. I built CiteTrace: provenance and attribution that shows which AI platform actually sent a visitor, so a brand holds a receipt instead of a vibe. Protocols get standardised and given away. Receipts get paid for.

And then the part I am proudest of, and the part that nearly broke me: forecasting. Monitoring tells a brand where it stands, and every competitor in this space stops there. The question a business actually pays for is different: if I do this specific thing, what happens to my visibility, and how sure are you? That is CitePulse, an action-conditional uplift forecasting engine built on a Bounded Confidence Model, with the calibration layer doing the honest work of making sure that when the system says 70 percent, it is right about 70 percent of the time. Calibrated uncertainty is unfashionable, unsexy, and the only defensible thing in this whole category.

None of that was fun in the way JD means fun. The fun parts fit in a weekend. The nine months were model updates invalidating baselines overnight, sampling budgets, timezone infrastructure decisions, and rewriting the extraction pipeline for the third time because a platform changed how it formats citations. Ideas are a spark. This was cardio.

And all of it happened in the gaps, because those nine months also contained a two year old and a three year old. The platform got built after bedtime, in the early mornings, on weekends traded like currency, with a laptop balanced somewhere between a nap schedule and a load of washing. Which is where I stop pretending any of this was a solo act. My wife carried our household through every late night and every distracted Saturday, and she is the actual superhero of this story. Founders love to talk about sacrifice as if it belongs to them. Most of it is quietly invoiced to the person standing next to us.

The call where it came full circle

Somewhere in the middle of those nine months, JD and I had another of our calls. Except this time, at the end, I shared my screen.

I showed him the dashboards. Live citation tracking across six platforms. Share of voice against named competitors. The attribution trail. The forecasts with their confidence bands. Months of daily data on brands that had no idea any of this was visible.

There was a pause, and then the man who wrote an entire essay about loving the part before the question dropped his tools and started building with me. He leads our data engineering now.

I want to be careful here, because this is the humble part and I mean it. I did not out-build JD. Read his piece: the man conjures working software out of specs for the joy of it, and Outercite’s codebase has his fingerprints all over it now. What I did was different in kind, not superior in skill. I stayed after the fun stopped. I picked the layer nobody wanted to build because it was boring, I aimed it at an industry rather than a plugin directory, and I kept grinding through the months where nothing was novel and everything was maintenance. He writes the better essay. I kept the receipts. It turns out a company needs both people, and the great luck of my year is that we ended up building the same one.

What you can steal from this

If you are sitting on your own repo, your own half-built thing, your own conversation that ended with a question you did not answer, here is the whole playbook, free, from someone who did it the slow way:

1. Your idea is not the asset. Your follow-through is. Someone else has your idea right now. Possibly at Google. The version that wins is the one attached to a person who refuses to stop.
2. Follow the second question. The bridge was my first idea. The meter was what building the bridge revealed. The valuable problem is usually one layer behind the one you started on, and you only find it by building the first thing.
3. Choose the boring layer. Everyone fights over the glamorous protocol. Nobody wins it. The measurement, the attribution, the receipt: that is where businesses quietly pay, forever.
4. Charge somebody early, and let it hurt. Both of us have shipped a version of the pricing page with nothing behind it. The braver move is the reverse: a real stranger, a real invoice, a product that embarrasses you. A stranger’s money is the only validation that cannot be argued with.
5. Find your complement. Builders need someone asking the yacht question. Askers of yacht questions need a builder who makes JSON come back correctly the first time. Neither of you is the whole company.
6. Grind is not a montage. It is nine months of unfashionable calibration work while the industry sells screenshots. It compounds anyway. That is the entire trick.

Where this is all heading

And now the part that gets me out of bed.

WebMCP will mature. The origin trials will end, the standard will settle, and the demand side, the agents themselves, will arrive, because every economic incentive points the same way. When that happens, the web stops being a library that machines skim and becomes a counter that machines transact across. Businesses will expose actions the way they expose opening hours. We registered for the origin trial on day one, because the thing I built a bridge toward in January is finally being paved.

But watch what happens after that, because it is bigger than websites.

Agents will not stay in browsers. They are already reaching into calendars, tills, and supply chains, and within a few years they will be walking around in hardware. When a humanoid robot collects a prescription, chooses a supplier, or books a service on behalf of its household, it will not see your signage, your branding, or your lovingly crafted homepage. It will consult whatever its model believes about you, weigh whoever its sources cite, and act. Every physical transaction will begin with an invisible act of machine trust.

Which means the question I fell into by accident, what do the machines actually say about you, and can you prove what it is worth, stops being a marketing question and becomes the commercial substrate of everything.

Visibility becomes trust. Trust becomes transactions. And every transaction, digital or physical, will need a receipt.

That is what we are building at Outercite. Not a dashboard. The source of truth for how machines see brands, and the proof of what that is worth, for a world where your next customer might not have a pulse.

JD ended his essay by saying something else was born from that conversation. This was it. Two repos, one question, nine months of grind, and a company.

The idea was free. We are building the receipts.

And the yacht remains, as ever, on the schedule.

If you want to know what the machines currently say about you, the free AI visibility check asks all six major AI engines a generic question in your category, never naming your business, and reports which of them recommended you. It takes about a minute and needs no account.

Arno Verburg is the founder of Outercite. Before this he built components for satellites and compliance software for pharmacies, which means he has spent his whole career making machines tell the truth. This piece is a follow-up to JD Conradie’s essay, When Google Validates Your Idea, which you should read first, INACTIVE labels and all.