What People Ask For When They Don't Know What to Watch
Asked in their own words, people name a mood or something to avoid about as often as a genre, and most requests name neither. This report sets out the figures, the method and the limits.
Among TasteRay requests that state a genre, a mood or something to avoid, 45.3% include a mood or avoidance. In free-typed opening messages that is 7.6%, against 9.2% for genre alone, and about two thirds match none of our categories. Over time, 34.3% of users stated a mood or avoidance. The sample is mostly Polish.
Drawn from more than 120,000 messages people sent TasteRay, from thousands of users, between January and November 2026. The figures below cover the more than 76,000 of those that were requests to Ray; answers to onboarding questions are left out. These are people using a recommendation product, mostly Polish users, so read this as evidence about what people type when given a box, and not as a survey of all viewers.
Mood, genre and everything else
About 80% of the messages were typed freely, and shares below use those as the base. "Opening" means the first message of a conversation, the closest thing to the first thing someone asks. Genre is the most common single constraint, and mood or avoidance runs close behind it. Most requests do something else entirely.
- 45.3%
- of requests naming a genre, a mood or an avoidance include a mood or an avoidance
- Opening messages. Across all messages the figure is 45.8%.
- 7.6%
- of free-typed opening requests state a mood or an avoidance
- This includes the 2.1% that also name a genre.
- 9.2%
- of free-typed opening requests name a genre with no mood or avoidance
- Words such as romantic and funny sit between genre and mood, so this boundary is soft.
- 34.3%
- of users stated a mood or an avoidance at least once
- Against 43.7% who named a genre. 10.9% stated a mood or avoidance and never named a genre.
- 67.8%
- of free-typed opening requests match none of our categories
- Lookups of a title by name, replies, generic asks and languages other than English and Polish.
- 8 words
- is the median length of a typed opening request
- About 45 characters. Across all typed messages, 7 words.
What people mention, by category
A message can carry several categories. The middle column is the share of free-typed opening requests. The right column is the share of users who mentioned the category at any point.
| Category | Share of opening requests | Share of users, at any point |
|---|---|---|
| Genre | 11.3% | 43.7% |
| Mood or feeling | 7.1% | 32.9% |
| Availability or platform | 6.4% | 25.7% |
| Similar to a named title | 4.4% | 24.3% |
| Occasion | 3.0% | 18.1% |
| Says they have already seen it | 2.0% | 15.1% |
| Question about a named title | 2.0% | 9.2% |
| Who is watching | 1.6% | 11.0% |
| Language preference | 1.6% | 9.4% |
| Avoidance | 0.6% | 6.2% |
| Time available | 0.6% | 3.9% |
The words people use for mood
Here are the mood concepts in rank order, by how many different people used each, across English and Polish. The top two are close, and so are several after them, so read the order loosely.
- Dark or gritty, then funny
- Sad or emotional, then light or easy
- Scary, then thought-provoking and romantic
- Cozy or comforting, weird or quirky, feel-good
- Further down: exciting, relaxing, tense, a pick-me-up, inspiring, mindless, nostalgic
English-speaking markets lean differently
Among accounts in the US, UK, Canada, Ireland, Australia and New Zealand, the order begins with funny, sad or emotional, scary and dark. Polish accounts lead with light or easy and dark. Mood or avoidance makes up 6.8% of free requests in the English-speaking group and 5.5% among Polish accounts. Questions about availability or a platform differ more, at 1.8% against 6.5%. The English-speaking segment is much smaller, and these are accounts of one product, so treat this as a hint worth testing on your own audience.
What people rule out
Avoidance in words is rare: 0.5% of free requests across all messages, from 6.2% of users. In rank order, the themes people named were scary titles or jump scares, animation or kids' content, cheesy or romantic material, slow or overly long titles, sex or nudity, and violence or gore. These only reach reportable size pooled across all messages. No theme did so in opening messages or within one country.
"I've already seen it" is a different signal: a reaction to a suggestion, in 2.0% of requests, from 15.1% of users.
When people ask
Sunday is the heaviest day, with 18.6% of free requests. The peak hours are 20:00 to 22:59, Warsaw time, at about 9% each. About 65% of requests arrive between 18:00 and 05:59. These hours fit Polish accounts and not other time zones, and nothing here describes how a streaming service's own traffic behaves.
What an operator can do with this
None of this proves a mood tab will work on your service. It does suggest where to start testing, at low cost.
- Name themes in the words people use
- Dark, funny, sad, light and scary are words people reach for unprompted. A first set of themes can start from them. Netflix's published examples "Something Light" and "Scare Me" are close to two of them, though its audience is not this one.
- Treat avoidance as a filter
- Few people type it, and those who do mean it. Offer a "not this" control, remove matching titles from your candidates, and send the rejection in
context.history. That needs mapping themes like jump scares or gore to your own metadata. - Filter out what they have seen
- More than one user in seven told Ray they had already seen a suggestion. Your service knows the watch history, so remove those titles before you rank anything.
- Answer availability from your side
- About a quarter of users asked at some point where something can be watched. On your own service the answer is simple, so show only what a viewer can play.
- Design for a short sentence
- A median of seven or eight words fits a single text box. Long forms ask for more than people volunteer. Suggestions help too: 21.6% of conversations opened with a preset or template, led by "What should I watch tonight?".
- Do not over-read the rare ones
- Time available appears for 3.9% of users, and that does not show a "Quick Watches" row would fail. People may simply not type it, and a button can draw out what a box does not.
- Prepare before the evening
- If your own logs show the same evening peak, refresh cached rows in the afternoon, since a response takes a few seconds and can run longer.
How we counted
- Source: the production database for Ray and the earlier TasteRay chat, 21 January to 5 November 2026. Answers to onboarding questions and WhatsApp were left out, and internal and test accounts were removed by email pattern.
- Each message was sorted into preset, acknowledgement, long (over 40 words, counted but not classified), empty or free-typed. Free-typed messages are the base for category shares.
- English and Polish word lists flagged each category. A negator followed by a theme within five words counts as avoidance, so "not too sad" is avoidance, not mood.
- Only aggregates left the analysis. Every phrase and cell reported here was used by at least 25 different people, and reference titles are never listed.
- Terms for mental health, grief, illness, trauma and relationship events are in no word list, so they are neither counted nor reported.
The limits
- These are TasteRay users, mostly Polish, who chose a recommendation product. They are not all viewers and not a representative sample of any country, nor people browsing a streaming service.
- About two thirds of requests are unclassified, in the 68% to 76% range depending on the view. Mood is probably undercounted, since many people express a feeling without a mood word.
- Classifier precision is roughly 55% to 80%: about 55% to 70% for mood, avoidance and who is watching, about 70% for time, about 80% for similarity. We checked small hand-labelled samples, with wide intervals of about 20 points either way, and did not re-sample after the last rule changes. Recall was not measured.
- Treat percentages as order of magnitude. Ranks and ratios are more trustworthy than exact figures.
- The product shapes what people type, because Ray offers suggestion chips, and those opening conversations are counted apart. Ray also covers other media than movies and TV series, and the word lists do not separate them.
- Request shares weight heavy users, which is why user shares sit beside them. Country is the account's stored country, and a team member on a private email address would not be caught by the exclusion.
Aggregated from anonymized requests. TasteRay is not affiliated with Netflix, whose theme names come from its post of 3 November 2026.
Sources and further reading
- Introducing AI Discover (Netflix, 3 November 2026)
FAQs about What People Ask For When They Don't Know What to Watch
Do people ask for a mood or a genre when choosing what to watch?
Both, about equally when they state either. In TasteRay requests that name a genre, a mood or something to avoid, 45.3% include a mood or avoidance. Most requests state neither, and the sample is mostly Polish users of a recommendation product, so treat this as direction and not as a fact about all viewers.
What mood words do people use most when asking what to watch?
In this report, the leading concepts were dark or gritty and funny, followed by sad or emotional, light or easy, and scary. Polish accounts led with light and dark, and English-speaking accounts led with funny and sad. Ranks are more trustworthy here than exact percentages, because the classifier is imprecise.
How often do viewers say what they want to avoid?
Rarely in words. Avoidance appears in 0.5% of free-typed TasteRay requests, from 6.2% of users at some point. The most common themes are scares, animation or kids' content, cheesy romance, slowness, sex or nudity, and violence. A filter control may draw out more than a text box does, which this data cannot show.
How reliable is this data?
It is useful as direction only. About two thirds of requests match no category, classifier precision is roughly 55% to 80% on small samples, and the people are mostly Polish users of one product, whose interface suggests prompts. Ranks and ratios hold up better than exact shares. The method and limits are on this page.