Users cannot evaluate content they never saw; absence produces no visible feedback

What it is

Dashboards have clicks, completions, complaints, unsubscribes. They do not have a column for “should have appeared and did not.” Users cannot file that ticket either — they do not know the title exists. Silent missing content means unshown items enter neither the user’s evaluation nor the system’s visible feedback. Coverage can shrink with no alarm on either the data side or the experience side.

That users cannot see what is missing is the perceptual layer. The further cut here: the product also receives no signal. In the optimisation loop, absence is zero, not negative.

Why it happens

Feedback channels register events that happened. Non-exposure is not an event; counters neither increment nor decrement. Complaints target visible bad, not unseen good. So “too narrow” is empty in the ticket system and in the loss; only positive events such as clicks speak. The objective already shrinks on its own; without a negative signal, the shrink can look healthy on the report.

Offline evaluation eats impressions and clicks that appeared in the log; absence is equally missing from the sample. An online A/B that watches only clicks can let both arms narrow in silence; the experiment will not flag a diversity failure. The alarm has to be bolted on: a contrast catalogue, an editorial floor, cross-user coverage — not a hope that users or the main metric will speak.

Studying it

Build a “should be visible” list outside the log (category quotas, editorial must-shows, random probe slots) and compare it to actual exposure. Dependent variables: share that should have been visible and was not, correlation of that share with complaint volume (expected near zero), change in evaluation after people are shown the missing list. Independent variables: whether probe slots exist, whether the report includes a miss rate.

Unknown-unknowns on the user side are a perception experiment. The method here is on the system side: make absence a countable object. If complaints and satisfaction stay flat while coverage falls, silence is working.

Where it stops holding

Search and an external book list make absence checkable, and silence breaks. Heavily edited media have internal meetings about “what did not make the page”; the alarm sits in people, not in users. When items live for hours, the “should be visible” list itself is unstable. This entry does not restate perceptual blindness, and it does not choose random versus structured injection.

Applying it

• Give coverage its own alarm: page if a must-keep class stays under quota for N days. Do not wait for clicks to fall.
• Keep a small stream of model-independent probe exposure, so items that exist but are never tapped still produce a visible record.
• Check: drop a must-keep class from the main sort for a week. Watch complaints, satisfaction, clicks. If the main metrics do not move and only the bolted-on miss rate does, absence will not speak for itself; the alarm has to be bolted on.

Related

• Same group: L6.08.1 Narrowing is produced by the objective itself; no one has to intend it · L6.08.2 Diversity costs short-term clicks, so it has to be an independent objective · L6.08.4 Random injection is not structured diversity; it only adds noise · L6.08.5 Diversity must sit on dimensions the user cares about; cross-category mix does not fix a single viewpoint
• Nearby: L6.02 Filter Bubbles · L6.13 Negative Feedback Channels for Recommendations · L6.09 Feedback Loops and Preference Entrenchment
• Search terms: silent missing content · no ticket for absence · coverage alerting