The one in four problem: why AI models pick your competitor

How it started

It started, like most useful things, by accident.

Last spring we were stress testing a new audit pipeline for a Copenhagen software company. The product was excellent. The reviews said so. The customers said so. The category analysts said so. So when we asked ChatGPT, Claude, Gemini, and Perplexity to recommend tools in that category, and our client showed up in one out of every four answers, the founder squinted at the screen for a long second.

"What about the other three?"

That question is the one we've spent the year answering for ourselves. A friend who runs comms at a Series B finance startup put it to us the same way in March. So did the head of growth at a healthcare brand with ten years of category dominance and a wall of awards. So did a craftsman with a one person workshop who wanted to know why a model recommended his competitor's bench when it had already described his work in glowing detail two messages earlier. Same question. Different scale. Same shape of answer.

From ten blue links to three names

Three years ago a B2B buyer searching for a tool got ten blue links. They skimmed. They formed their own ranking. They argued about it with a colleague over coffee. Buyers brought their own judgment to the page, and the page was a long page.

Today that buyer asks ChatGPT. Or Claude. Or Perplexity. They get three names. Not ten. Not a sorted list to negotiate with. Three confident recommendations delivered like a friend whispering an answer.

Three names. Then a decision.

The implications, when you sit with them, are uncomfortable. The model is doing the shortlist. Not the buyer. The buyer is doing the contracts, the calls, the demo. The shortlist, the thing marketing teams used to fight tooth and nail for, is happening earlier and somewhere a brand doesn't get to see.

What is signal

What's strange is that the recommended brands are not always the best products. We've watched this play out across many categories now. The leader by revenue gets recommended, sometimes. The leader by review score gets recommended, sometimes. The leader by signal gets recommended almost every time.

A model recommending you doesn't know you. It hasn't used your product. It hasn't met you. What it has is a corpus of language: things you've published, things customers have written about you, things journalists have written about you, structured metadata and the way third parties speak your name into the world. When the model considers a question like "what is the best X for Y," it surveys that corpus in its own quiet way and looks for brands whose presence in the corpus paints a recognizable picture against the question.

Signal is the shape of the picture your published language paints. It is what the corpus collectively says about you.

Why two great products get different recommendations

Two brands can have identical product quality and very different signal. We've seen this many times now, often within the same category. The winner has consistent customer language, a clear structured description of what the product is for, coverage from outside the company that triangulates with the company's own voice, and a category vocabulary the model can map onto. The loser has scattered language, an absent middle layer of explanation, and a heavy reliance on phrases their competitors use too.

When a model meets that loser, it doesn't reject them. It just doesn't reach for them.

You can see this directly. Open ChatGPT today, ask it to recommend three companies in your category, and watch what it does. Watch the language it uses. Watch which names come back first, second, third. Watch which names don't appear at all even though you would expect them. The asymmetry between brand A and brand B in the same room is sometimes huge. The product quality gap between brand A and brand B is sometimes nothing.

What we publish here

So that's the one in four problem. Most brands are getting recommended in one out of every four answers in their space. The leaders are getting recommended in three. The gap will widen as buyers route more decisions through models, which they will, because the alternative is reading ten blue links and arguing over coffee, and people have started to find that tiresome.

The pleasant news is that signal is a thing you can work on. The frustrating news is that nobody has, before now, told you exactly which signal to work on, what your competitors are doing that you aren't, and which fix matters most.

Signal is the audit. It runs live ChatGPT and Gemini queries against your real category, scores you across five pillars, benchmarks you against three competitors, and produces a 10 to 15 page PDF that ends in an action plan. Every finding gets walked through with you on a 30 minute call within one business day. We've sent enough of these out by now that we're starting to recognize patterns we want to write about, and that is the purpose of this journal.

Some of what we publish here will be diagnostic: this is what good signal looks like, this is what poor signal looks like, here is how the four big models differ in how they weight it. Some will be tactical: this is the cheapest fix we have seen move a score, this is the most expensive one, this is the one nobody bothers with that pays the most. Some will be opinionated. We'll try to keep the opinionated ones honest.

Next week we'll walk through the four signal characteristics shared by the brands that come up first in their category. They are not what most marketing teams expect.

If you would rather see your own score first, run Signal. It is €690 for the report and the walkthrough.