Clickbait headlines
A Socratic walk-through of clickbait headlines — reasoned out one step at a time, not lectured.
The question we started with
THE QUESTION #Why do headlines grow more tempting and less informative as publishers learn what readers click?
A newspaper headline of the old kind tried to tell you the story: who did what, and how much. A headline optimised for clicks tries to make you need to find out — you won't believe what happened next, this one detail changes everything. It withholds the very thing the old headline supplied.
The strange part is that this is the result of learning. Publishers now measure which headlines get clicked, with far better data than any editor ever had, and feed that back into how headlines are written. Better information about readers has produced headlines that serve readers worse. That is worth explaining, because normally we expect measurement to improve things.
Reasoning it through
REASONING #Start with what a headline is doing. It is a summary, and it is also an advertisement for the article. Those two functions used to be nearly aligned: the best way to persuade someone to read a story was to tell them what was in it, because a reader who wanted that story would go and read it.
Now ask what happens when you can measure clicks, and only clicks.
Consider two headlines for the same article. One states the finding. The other implies there is a finding and does not say what it is. Which gets more clicks? The second, reliably — because the first one already gave the reader what they wanted. A reader who learns from the headline that a council raised parking charges by ten per cent may be perfectly satisfied and move on. That is a successful act of communication and a failed act of monetisation.
That is the core of it. Informativeness and click-through are in direct tension, because information delivered in the headline is information the reader no longer needs to click for. The headline that serves the reader best is the one that most often makes clicking unnecessary.
Now add the feedback loop. Test two headlines, keep the winner, repeat. Nothing in that loop measures whether the reader was glad they clicked, whether they read to the end, whether they understood the story, or whether they trust the publication next week. It measures the click, because the click is what can be counted immediately and what is sold to advertisers. So the loop hill-climbs on a proxy, and it climbs efficiently — which means it moves fast toward whatever maximises the proxy at the expense of everything the proxy fails to capture.
That is why more measurement made things worse rather than better. A weak measurement of a good thing was replaced by a precise measurement of a proxy, and precision applied to a proxy is exactly what drives the proxy and the goal apart. This is Goodhart's problem in its ordinary form, and the "curiosity gap" headline is the shape the optimum takes in this particular space.
There is one more term, and it is what makes the outcome collective rather than individual. A publisher who declines to write this way loses traffic now to publishers who do, in a shared feed where everyone's headlines sit side by side. The reader's attention is the contested resource, and restraint is unilateral disarmament. So even a publisher who believes the practice erodes trust has to weigh a certain loss today against a diffuse cost later.
The analogy
THE ANALOGY #Think of a shop that measures how many people walk through the door and pays its window-dresser on that number alone.
At first the window shows the goods, and people who want them come in. Then the dresser discovers that a covered shape with a sign saying the thing everyone is talking about brings more people through the door than a shelf of visible merchandise. The measured number goes up. Some of those people buy nothing and leave irritated, but that is not in the number.
Run this for a year and the window no longer tells anyone what the shop sells. The dresser has not failed; they have succeeded completely at the task they were set.
A shop's customers can see the goods once inside and judge for themselves, whereas a reader's disappointment is diffuse and rarely traced back to the headline that caused it — so the corrective feedback that would eventually punish the shop arrives, if at all, much later and much weaker in the publishing case.
Clarifying the model
THE MODEL #This is not a claim that headline writers are cynical, and the mechanism does not need them to be. The loop works through A/B testing and traffic reports, not through anyone deciding to mislead. A writer producing what performs is doing their job as defined. Locating the cause in individual character rather than in the measurement is the standard error here, and it leads to remedies — exhortation, style guides, professional shaming — that do not touch the incentive.
Structural cause versus technological affordance. It is tempting to blame the technology: the internet made this happen. But the affordance and the incentive are separable, and it is the incentive doing the work. Headlines were sensational in the era of street sales too, when revenue depended on the copy being picked up from a stand and the visible headline was the entire pitch. Subscription-funded publications with the same technology behave differently, because the thing being sold is a relationship rather than a click. The technology supplied precision; the funding model supplied the direction.
The equilibrium is unstable in a way that matters. Readers learn. A curiosity-gap headline that reliably disappoints trains people to ignore that form, which is why the specific formulations churn — each one works until it is recognised, then decays. So the practice is not a stable optimum but a treadmill, and the observable churn of headline styles is decent evidence for the account.
Platforms have partially corrected, which is the natural experiment. When intermediaries began weighting dwell time, scroll depth and returns rather than raw clicks, the payoff to pure curiosity-gap headlines fell and the style became less dominant in those channels. That is the prediction the account makes: change what is counted and the writing changes, without anyone's values changing at all.
What I am unsure of. How much of the drift is publisher optimisation and how much is competitive selection — outlets that happened to write this way surviving while others failed — is not something I can separate. Both mechanisms predict the same drift.
The falsification test. If the driver is optimisation against a click-only proxy, then publications funded by subscription or by measured reading time should show measurably more informative headlines for the same stories than advertising-funded ones optimising on clicks. If headline informativeness were the same across funding models, the incentive account would be wrong and something about the medium or about readers' preferences would have to explain it instead.
A picture of it
THE PICTURE #How to readThe horizontal axis is how much the headline gives away; the vertical is how well it performs on the metric being optimised. Placements express the argument rather than measured data. The important feature is which quadrant is nearly empty: headlines that both state the finding and out-click everything else are rare, which is why the optimiser drifts left. The one point placed in that quadrant is the case where the finding is itself surprising enough to sell the story — the exception that shows the tension is a tendency, not a law.
What became clearer
WHAT CLEARED #Headlines did not get worse because writers got worse or readers got dumber. They got worse because a precise measurement was introduced for something that is not the goal, and the thing measured is in direct tension with the thing wanted: any information the headline supplies is information the reader no longer has to click for. Optimising hard against that proxy necessarily moves away from informativeness — and the fix is not better writers but a different quantity to count.
Where to go next
ONWARD #- Why dwell-time and return-visit metrics change the writing without changing anyone's values.
- How the churn of headline formulas is evidence that readers adapt to each one.
- Why subscription funding pulls in the opposite direction, and what it costs in reach.