Search and history that are not personalized must exist as a fallback
Aliases: unpersonalised search · watch history · escape from the feed
What it is
Coordinates will not stick, yesterday’s row will not come back, screenshots will not match — those failures still have one retreat: search by title, open watch history. If search’s default is still re-ranked “for you,” and history is folded into the feed by personalization, the retreat breaks too. Non-personalized fallback paths means search and history must be able to work in a mode that does not eat a personal vector, as addressing channels when spatial and temporal memory fail.
Fallback is not turning personalization off everywhere. It is two entries that “best for now” is not allowed to rewrite.
Why it happens
Addressing has at least three layers: coordinates, time, identifiers. When personalization wounds the first two, the third — names, IDs, records the user left — must still work. If search multiplies a personal vector into ranking, same-named items surface by taste rather than by match, and the identifier layer is polluted. If history only shows what the model thinks you will watch again, yesterday’s flicked-past row stays unreachable.
Fallback also carries the interpersonal job: support says “search the title,” a friend sends a name, and both people can meet on the same result. Without it, a non-shareable personalised UI becomes unrepairable.
Studying it
Leave content personalization as it is; vary only search and history: search with a personal vector versus pure match; history of everything that appeared versus only what the model kept. Tasks: re-find yesterday’s unfinished item, meet on a title a friend named. Dependent variables: success, whether people wander into For You, judgment of “is search guessing me too.” Independent variables: discoverability of the entries, whether they default to personalised.
Do not measure re-finding on the personalised home as a substitute for this question. That measures the stream. Here the question is whether a second addressing system exists after the stream fails.
Where it stops holding
Spelling correction and synonym expansion in search are not personalization and can stay. Continue-watching that ranks by account rather than by taste counts as history, not personalization. Enterprise search has permission filters; that is authorisation, not taste. This entry wants the fallback to exist and to be usable by default. It does not restate stable structure, and it does not require turning personalization off for the whole product.
Applying it
- Search defaults to match quality; the personal vector is an optional layer, not the only sort. History lists what appeared and what was watched, by time, not hidden by “you probably do not want this again.”
- Both entries have a fixed seat on the home skeleton and are not hidden per person.
- Check: with a deliberately skewed profile, search a rare title just seen on home, and see whether it still sits on the first screen of results. Then use history to find an unfinished item the model would not push again. If both fail, the fallback does not exist.
Related
- Same group: L6.12.1 Personalization makes the same entry show different content, so location memory fails · L6.12.2 Content that changes over time makes previously seen items hard to find again · L6.12.3 Users cannot give each other directions; shared screenshots no longer match anyone else’s interface · L6.12.4 Stable structure with variable content is easier to master than a structure that also moves
- Nearby: L6.05 Turning Personalization Off · L6.10 Turning Personalization Off and Resetting It · L2.07 Prompt History and Reuse
- Search terms:
non-personalized fallback paths·unpersonalised search·watch history
Cards in the same group
- L6.12.1Personalization makes the same entry show different content, so location memory fails
- L6.12.2Content that changes over time makes previously seen items hard to find again
- L6.12.3Users cannot give each other directions; shared screenshots no longer match anyone else’s interface
- L6.12.4Stable structure with variable content is easier to master than a structure that also moves