If everything is an algorithm, why not build one to fight it?

Let me show you a Monday. Simple objective, one client.

I need to know what to write next. So I open Search Console first, and I go looking for impression gaps. Queries where we’re being seen and not clicked. Countries where we’re showing up and shouldn’t be, or should and aren’t.

But I don’t look at seven days.

That’s the part most people get wrong, and it took me years of losing traffic to learn it. Seven days tells you about last week. So I pull thirty days, then three months, then six. And I’m not only looking at what went up. I’m looking at what came down, alongside the queries that came down with it.

Because a good algorithm knows what not to do.

That’s the bit nobody builds. Anyone can find the winners. What’s harder is recognising a pattern of losses you’ve seen before. Was the December 2025 core update shaped like the June 2022 one? If a cluster of queries is sliding, did a similar cluster slide the same way three years ago? And what turned out to be underneath it?

I can’t see Google’s signals. Nobody can. But I’ve watched enough things break to infer which ones are moving. The inference comes from the failures, not the wins.

Then I go and get what the machine can’t tell me. I’m in touch with sales, because they hear the question phrased the way a buyer phrases it. I’m in touch with product, because they know what’s shipping. And I keep a competitive deck that updates itself with AI. Alongside SEMrush and Ahrefs, wired in through their MCP protocols. So I know what the other side moved on this week without going to look.

All of that comes together and produces the next five topics.

An algorithm is a series of processes. Read that list again.

Now go back over what I just described.

Five sources. A fixed order. Rules for what to pull from each. Time windows I always use. A comparison step. A pattern-matching step against things that already happened.

That’s not a workflow. That’s an algorithm. An algorithm is a series of processes that include decision rules. I could hand it to somebody tomorrow and they’d arrive at roughly the same five topics.

Call it eighty percent. Eighty percent of how I decide what to write is already deterministic enough to write down.

And here’s the uncomfortable part. The last twenty percent isn’t as special as I’d like it to be either. What to write. When to publish it. What to link, and where. Because I have steps for that too. I follow them. They’re just less examined. Because I have steps for that too. I follow them. They’re just less examined.

So we’ve got a person running an algorithm, pointed at a company running an algorithm. With a set of signals I can never see. Signals I infer from fifteen years of watching things break.

Which makes the question in the title a fair one rather than a provocation. If everything here is an algorithm, why not build one to fight it?

I don’t have a clever objection. What I have is somebody else’s answer.

So let me start with the version that has already happened.

High-frequency trading is the purest form of this idea anyone has attempted. Algorithms fighting algorithms, with everything else stripped away.

And notice how much better the conditions were there than they are in your job. Outcomes measured in microseconds. No brand to protect. No tone of voice. No legal review, no quarterly narrative, no client who wants the logo bigger.

Capital was effectively unlimited. The people were physicists. And the objective function fit on a napkin. Be first.

So if building a machine to beat a machine works anywhere, I’d expect it to work there. Let’s see what it bought them.

The window kept closing.

Budish, Cramton, and Shim measured it precisely in the 2015 Quarterly Journal of Economics paper that named the problem.

Let’s look at the one arbitrage they picked. The E-mini S&P 500 future trades in Chicago. The SPDR S&P 500 ETF trades in New York. They track the same thing. Over an hour they’re perfectly correlated. Over a millisecond the correlation collapses entirely, which leaves a gap, and the gap belongs to whoever gets there first.

Median duration of that gap in 2005: 97 milliseconds.

Median duration in 2011: 7 milliseconds.

Six years, and the space where the money lived had shrunk by more than an order of magnitude. Every improvement anybody made became the baseline everybody else had to match within months.

Now, the spending went the other way. Firms laid private fibre. Then somebody worked out that microwave beats glass over land. So they started putting up relay towers across Illinois farmland, to shave off the difference. The exchanges saw where this was going and began selling co-location. That’s a rack a few metres closer to the matching engine. Priced according to how badly you needed those metres.

So the window shrank and the bill grew, which brings us to the number we came for.

Coz the prize never moved.

Over those six years, spending kept rising steadily. But the profitability of each arbitrage stayed remarkably constant. A median of about 0.08 index points per unit traded.

It didn’t shrink; it didn’t grow. It sat there while we all watched.

I sat with that number for a while, because a shrinking one would have let everyone off the hook. You could call that efficiency. Competition eroding an inefficiency, exactly as the textbook says.

That isn’t what happened. Hundreds of millions of dollars went into speed. The prize was the same size at the end as it was at the start. The authors’ conclusion is that the arms race doesn’t change the size of the prize at all. It just keeps raising the bar for how fast you have to be to reach it.

It’s the shape rather than the numbers that I keep coming back to. A constant prize. A rising cost of entry. A shrinking window. And a handful of firms taking most of what’s left.

That’s an arms race in the technical sense, not the loose one. Only relative position matters. So everyone keeps spending to hold the place they already had, and nobody can stop on their own.

And the clearest evidence of who won isn’t in the trading profits. It’s in where the money settled, which was with the people selling cable, towers and rack space.

In an arms race, the arms dealers do fine.

Your version is worse, and here’s the part that stings.

If that felt flattering to your side of the fence, I don’t think it should.

The finance version had one property yours doesn’t have. It was symmetric. Everyone fought everyone under one immutable constraint. The speed of light. No exchange was ever going to revise that on a Tuesday afternoon.

Your adversary owns the track.

The platform you’re modelling can change its function whenever it likes. It has a direct financial interest in your strategy failing. And it sees your output sitting alongside thousands of other attempts at exactly the same thing.

That last point gets underrated. The corpus of everyone trying to game the system is the training data for the classifier that catches you.

We’re not racing a rival. We’re racing the referee, and the referee reads your playbook between matches.

And then there is a deeper problem, which has nothing to do with getting caught.

Most of us stop at that objection and treat it as risk management. Hedge it, diversify, keep a second channel warm.

There’s a worse one underneath, and I found it in an odd place.

Biology and economics arrived at the same principle from opposite ends in the 1970s. In economics it came through job market signalling. In biology it came through the question of why peacocks grow tails that make them easier to eat.

Both landed on the same answer, and I’d stake this whole piece on it. A signal is reliable only when faking it costs more than the lie could earn. Reliability comes from differential cost. Expensive for a faker, affordable for the real thing.

Now let’s apply that to every marketing signal that has ever worked.

A backlink meant something because getting one meant convincing an editor who could refuse you. A review meant something because it required somebody who bought the thing. Used it. And formed an opinion strong enough to write down. A community mention meant something because a stranger typed your name for nothing.

In each case, faking it cost more than earning it. That gap was the entire information content of the signal.

Automation closes the gap. Which is precisely what automation is for.

The moment producing a signal becomes free, it stops telling the receiving system anything. And that forces a response that has nothing to do with policy or fairness or enforcement. A system that keeps weighting a corrupted signal starts returning worse answers. A retrieval system that returns worse answers loses its users to one that doesn’t.

Reweighting isn’t punishment; it’s survival.

So the consequence follows directly. The better you automate a signal, the faster you destroy the reason anyone was looking at it.

You’re not extracting value from the system. You’re consuming it. And there’s less left in every cycle, including for you.

Every channel we’ve ever called saturated went through exactly this. Saturation was never about volume; it was about the signal ceasing to mean anything.

Hence, a million-dollar question. Should you automate, or not?

Go and look at how this gets answered in public and you’ll find one debate, repeated everywhere. Will Google penalise me for AI content?

The answer is always the same. No. Google penalises low quality, not AI. Google’s own guidance has said exactly that for three years, and every practitioner writing about it says it back. (If you believe them.)

That’s true. Yes.

It’s also useless. (Sorry.)

Because it answers a question about enforcement, and enforcement was never the risk. Go back to what we established a few paragraphs ago. A system that keeps weighting a corrupted signal starts returning worse answers. A system returning worse answers loses its users to one that doesn’t.

So the signal dies whether or not anybody catches you. Nobody has to enforce a thing. The tactic stops working because it stopped carrying information. That happens quietly, on its own schedule. No penalty notice, and nothing to appeal.

The whole industry is watching the wrong dial.

One objection survives everything above, and I can’t get round it. Unwinnable in aggregate is not the same as unprofitable for you.

The trading arms race wiped out a lot of value in the industry. At the same time, it made some firms very wealthy. Being fastest in a race nobody can win still beats being slow in the same race. The equilibrium is terrible collectively and excellent for whoever happens to be ahead.

So anyone advising you against automation on systems-level grounds is handing you advice. Advice that would have lost money for fifteen consecutive years.

Which leaves me owing you an actual answer. Here it is, as a rule rather than a principle.

Automate anything whose value doesn’t depend on it being expensive to produce.

That covers most of what you do. Reporting. Data cleaning. Keyword clustering. Translation. Ad variants. QA. Brief generation. Migration checks. Nobody on the receiving end reads any of it as a signal. So nothing degrades when the cost goes to zero. The value was always the work you didn’t have to do.

Automate all of it. Permanently.

And stop feeling clever about it, because this is the boring half and it’s still mostly undone. Then there’s the other kind, and here’s where I’ll disagree with the people telling you to stay away from it.

Anything whose value depends on being costly still works. The mention, the review, the link, the page that exists to look like effort. It works today. It’ll keep working for a while. And every unit you produce spends down the thing that made it worth producing.

That’s not a reason to avoid it. It’s a reason to price it as a trade rather than an investment.

Take the money. Know it decays. Don’t build the company on top of it. The firms that got rich in the arms race knew they were renting an edge. They kept a second business running. The ones who thought they owned it are the ones still paying for towers.

So that’s the answer. Automate the boring half without guilt, forever. Automate the manufactured half with an exit date written down before you start.

And that’s what fifteen years of the fastest arms race in history bought. Not an edge, but a tax paid to the people selling the equipment in exchange for standing still.

Point automation at production and you’re in that race against an opponent who owns the rulebook. Running a strategy that erodes its own returns in proportion to how well it works.

Point it at observation and you’re only competing. The edge still gets competed away, the way every edge does. But the thing you were looking at is still there afterwards.

One race consumes what it runs on. The other doesn’t.

Which takes me back to my Monday. Every one of those five sources is observation. Impression gaps, six-month movement, what came down and when, what the competitors moved on, what sales is hearing. None of it manufactures a signal. All of it reads one.

And almost none of it is stitched together anywhere. I run it across four tools, a deck, and two conversations. And I hold the pattern-matching in my head, because there’s nowhere to put it.

That’s the thing worth building. Not a machine that writes. A machine that watches, remembers what broke last time, and tells you when this looks like that.

So let’s automate the seeing, and do the rest ourselves.