In 2021, I was running a venture-backed analytics platform in a market that was going vertical. Usage was climbing, the numbers looked excellent, and we raised on them. I read that growth as validation of the product.
It wasn't. It was validation of the market's temperature. When the market cooled, most of it went with it, and what remained was the actual demand, which had been there the whole time underneath the noise and was a much smaller and much more informative number.
That mistake cost more than any technical decision I have made, and I've since come to think it is the most common analytical error in fast-moving sectors. It's worth being precise about the mechanism.
Why the two are so easily confused
Both produce the same observable: people showing up.
Attention is a response to salience. Something is being discussed, so people investigate. The behavior looks identical to demand at the top of the funnel and it's driven by an entirely different thing: what's in the discourse this month.
Demand is a response to a problem. Someone has a recurring need and your thing addresses it. The tell isn't volume, it's what happens when the noise stops.
The reason this is hard is that during the rise, both curves point the same direction, and every dashboard you have is showing you the sum. There is no metric that separates them in real time. You find out afterward, which is exactly when it stops being useful.
The distinguishing signals, in advance
There are some, and I didn't look at any of them at the time.
Return behavior under quiet conditions. Not retention in aggregate, which stays flat while new attention-driven users replace the ones leaving. The specific question is: what does a cohort that arrived in a quiet week do, compared to a cohort that arrived in a loud week? If the quiet cohort behaves better, you have demand with attention on top. If they behave the same, you may have nothing but attention.
Whether usage survives a competitor's launch. Attention is zero-sum and moves to whatever is newest. Demand isn't. A user with a real problem doesn't leave because something shinier appeared, they leave because something solved the problem better. Watching what happens to your numbers when a competitor gets a lot of coverage is a cheap and sharp test.
What people say when you ask what they would use instead. A demand-driven user names a specific alternative and describes why it is worse. An attention-driven user says they would probably just stop. That second answer is the whole finding, and one afternoon of conversations surfaces it.
Whether anyone is paying for the boring part. In almost every market there's an unglamorous component that only people with a real problem care about. Exports, reliability guarantees, integration with something ugly, support. Attention never buys these. If nobody is asking for the boring part, nobody has the problem badly.
The macro version of the same error
This scales up beyond a single company. Sector-level analysis mistakes attention for demand constantly, and the mechanism is identical.
A category gets discussed heavily. Capital flows in on the basis of engagement metrics. Those metrics are real and they are measuring interest. Then interest moves elsewhere, the metrics fall, and everyone concludes the category failed, when what actually happened is that the underlying demand was always a fraction of the observed activity and is now visible for the first time.
Notably, the underlying demand is often fine. It's just an order of magnitude smaller than the peak suggested, which means the businesses built for the peak do not survive to serve it. That's not a failure of the thesis. It is a failure of sizing, caused by measuring the wrong quantity.
What I do differently
I now try to estimate the quiet-period baseline explicitly rather than extrapolating from the trend. The question I ask is: if nobody talked about this category for a year, what would still be happening?
That number is unpleasant to produce and always lower than the current one. It's also the only number worth planning against, because everything above it is borrowed from a discourse you don't control.
I would add one caution against overcorrecting. Attention is not worthless. It's a genuine, temporary resource, and using it to acquire users who turn out to have real demand is legitimate and effective. The error isn't benefiting from it. The error is putting it in the denominator of your growth model and building a cost structure on top.
Treat attention as weather. Plan for the climate.