The recommendation panels at the foot of articles are running an optimisation process in public. Understanding what they optimise for explains why they look the way they do.
Every headline is competing continuously
Advertisers supply multiple headline and image combinations for the same destination, and the system distributes impressions across them to observe response.
Combinations that attract clicks receive progressively more of the available inventory, while weaker ones are throttled within hours rather than days.
The visible result is therefore the survivor of a large number of comparisons, which is why these units converge on similar phrasing across unrelated advertisers.
Curiosity outperforms description
Headlines that state the conclusion give the reader everything they need without a click, and they lose consistently to headlines that withhold it.
The selection process has no preference for one style over another. It simply promotes whatever produced the behaviour it was measuring.
The characteristic tone of these units is thus an output of the optimisation rather than a stylistic decision made by anyone in particular.
Placement context is part of the calculation
The same combination performs differently depending on the article it appears beneath and the audience that publisher attracts.
Systems account for this by learning per-placement, which means an advertiser's best performer on one site may be suppressed on another.
This is why widget performance is difficult to forecast from a small test. The unit of learning is the pairing of creative and placement, not the creative alone.
Post-click behaviour has become an input
Optimising purely for clicks rewards headlines that overstate, and the resulting traffic leaves immediately, which damages the publisher's experience.
Networks responded by incorporating downstream signals such as time on the destination and repeat visitation into what they promote.
The effect is partial rather than complete, since click behaviour remains abundant and immediate while downstream signals are sparse and delayed.
Image selection often outweighs wording
In a small unit competing for peripheral attention, the thumbnail is processed before the text and does most of the work of stopping the eye.
Testing images against a fixed headline usually produces larger swings than testing headlines against a fixed image, which surprises advertisers who arrive from search backgrounds.
The practical implication is that image variety should be the first axis of testing, and headline refinement the second rather than the reverse.