A single dimension drives one-sided optimization

What it is

When a measurement system makes only one dimension the visible target, resources, experiments, and interface levers concentrate on the knobs that move that dimension. Other dimensions can worsen in parallel without entering the decision. That is single-dimension local optimization. It is not merely “the wrong number.” After the taxonomy is cut down, the organization has only one rewarded direction. If daily actives are the only axis, notifications, badges, and autoplay win; whether the task completed or the person felt interrupted stops counting.

Why it happens

Almost every interface change touches several dimensions, but review pays only the counted column. Teams learn that levers which move that column are “effective design,” and that the rest is talk. Engagement can be pushed quickly through session length or launch counts; happiness and task success move more slowly and more noisily, so they lose attention inside the same iteration cadence. One-sidedness is not the absence of other dimensions; it is their absence from the objective. When a change lifts the single dimension and drops a neighbor, the drop is labeled noise or “later,” and the rise is taken as proof the design was right. After a few cycles the product morphology grows around that local optimum, and recovering the other axes means tearing out levers that have already grown in.

Studying it

Back-read version history on multiple axes: find intervals where the sole rewarded dimension rose steadily, and plot the others for a stable inverse move. At the experiment layer, compare a “optimize one axis” shipping rule with a multi-axis constraint over the same period, scoring net change on the unrewarded dimensions rather than the headline metric alone. Decision-trace analysis also works: the narrower the metric set cited in review notes, the more subsequent launches cluster on that dimension’s short levers. On the qualitative side, ask teams what would count as a successful change, and code whether answers map to a single axis.

Where it stops holding

A deliberate, time-boxed focus on one dimension can be a stage choice, provided other dimensions are labeled “not this period, but not to be breached.” If those dimensions are not collected, breach cannot be judged. The harm looks different in growth and monetization: the first often trades interruption for activity, the second trades friction for conversion; the mechanism is the same. Collapsing every dimension into a new weighted “single metric” only hides the one-sided drive; it does not remove it.

Applying it

• Name the currently rewarded dimension and the three interface levers that move it fastest.
• For each lever, pre-write a column of dimensions that could be sacrificed; experiments without that column do not ship to test.
• If the headline rises and any undeclared dimension falls, mark the result as a fail, not as a win on the headline.
• Each quarter, inspect the lever mix of recent launches; if all sit on one dimension’s short levers, require the next round to serve a different axis.

Related

• Same group: Q6.01.1 Measure across happiness, engagement, adoption, retention, and task success · Q6.01.3 Not every product needs every dimension
• Adjacent: Q6.03 North star metrics · Q6.05 Metric manipulability
• Search terms: single-dimension local optimization · one-sided optimization · metric myopia