Demand is sampled median video performance: median views-per-day among relevant videos returned for this keyword in the last 30 days, split by length. Competition is estimated, not an exact count: the API estimate is scaled by the on-topic share observed in the same format. Supply movement applies that identical relevance rule to both recent and comparison cohorts. Relative rank orders the eligible keywords within its own format: a Short is scored against other Shorts, never against long form, because the two sit orders of magnitude apart on views per day. Rank is descriptive, not a mark out of 100; absolute demand, sample, peer, channel-spread and stability gates control the verdict. The verdict is a content-production call for today, not a recommendation to remove the keyword from the locked evidence panel. Keywords whose results were mostly off-topic are excluded rather than scored.

6 keyword/format row(s) could not be ranked. They are omitted from the table rather than shown with a guessed score.

• pcos 7 day meal plan long-form: unranked: only 20% of results were on-topic, the search term is too ambiguous to measure (3 of 15 sampled videos were on-topic, 20%. Try a more specific search term.)
• pcos dessert recipes long-form: unranked: only 0% of results were on-topic, the search term is too ambiguous to measure (0 of 2 sampled videos were on-topic, 0%. Try a more specific search term.)
• pcos high protein breakfast long-form: unranked: only 20% of comparison results were on-topic, so supply movement is not like-for-like (Comparison window: 1 of 5 sampled videos were on-topic, 20%. Recent window: 4 of 7 on-topic, 57%.)
• pmos dessert recipes long-form: unranked: only 0% of results were on-topic, the search term is too ambiguous to measure (0 of 2 sampled videos were on-topic, 0%. Try a more specific search term.)
• pmos diet long-form: unranked: only 38% of results were on-topic, the search term is too ambiguous to measure (6 of 16 sampled videos were on-topic, 38%. Try a more specific search term.)
• pmos high protein breakfast long-form: unranked: only 25% of results were on-topic, the search term is too ambiguous to measure (1 of 4 sampled videos were on-topic, 25%. Try a more specific search term.)

Highest observed videos in the relevant sample. Up to six per format are shown after collapsing near-identical titles and limiting any one channel to two rows. They are observations from YouTube's relevance-ranked sample, not proof of what universally "works". Every row links to its source and names the panel searches that found it.

Title words observed in relevant samples. This is a correlation, not a cause. Words and sampled median video performance are calculated separately for Shorts and long form. Nothing here shows that a word caused the views; the sample is relevance-ranked, not random. Treat these as format-specific prompts, not rules.

Written against the keywords that scored highest above. The evidence line under each is copied from the measured data, not written by the model.

Stated up front rather than implied by omission. Every item here is a real limitation of the data behind this report.

• Demand momentum is unavailable: there is no accepted completed observation at the configured 7-14 day separation. A repeat run on the same day is deliberately not treated as a trend baseline.
• 5 keyword/format pair(s) could not support a demand reading and are unranked: pcos 7 day meal plan (long), pcos dessert recipes (long), pmos dessert recipes (long), pmos diet (long), pmos high protein breakfast (long). The row-specific reasons distinguish too few relevant videos, low relevance, genuine zero, and incomplete evidence. A keyword can rank in one length and not the other; that is a real finding about where the topic lives, not a gap in the data.
• 6 keyword/format pair(s) have no comparable supply-momentum ratio because evidence was incomplete, relevance was insufficient, or the comparison denominator was a genuine zero: pcos 7 day meal plan (long), pcos dessert recipes (long), pcos high protein breakfast (long), pmos dessert recipes (long), pmos diet (long), pmos high protein breakfast (long).
• The short-form cohort is identified by recorded runtime of three minutes or less, because the Data API exposes no Shorts flag and no vertical/aspect field. Checked against YouTube's own classification on 150 real videos, that cutoff agreed 97.3% of the time: 4 wrong out of 150, none missed, and every error was a brief regular video assigned to that cohort. The check therefore observed a few percent of contamination in the short-form cohort and none in the long-form sample; it does not prove zero error outside that check. Videos whose runtime the API did not report, often live broadcasts, are counted in neither.
• Competing supply is reported per length, but the API only estimates a total for the keyword as a whole. The split between the two lengths is the share each took of the sample, applied to that estimate. It is an approximation laid on an approximation, and is the softest number in this report.
• Demand momentum compares the last-30-day cohort against the same cohort measured on an earlier run. Consecutive days overlap by 29 of 30 days, so day-over-day movement is heavily damped by design. Week-over-week change is the meaningful reading; a single day's shift is not.
• A keyword is only called 'build' if its on-topic long-form videos take a median of at least 20 views/day. For the short-form cohort the bar is 200 views/day instead, because the two lengths do not share a view scale and one floor for both would pass essentially every short-form row. A lower Test call still requires 5 long-form views/day and 50 short-form views/day. Today 19 of 20 short-form readings sat under it, and 14 of 15 long-form readings sat under it. Relative rank necessarily has a winner even in a uniformly weak set, so every positive call also needs enough absolute demand, relevant videos, peer keywords and channel spread, and must survive removal of its fastest video. These floors are judgement calls set in the niche config, not measured constants.
• Supply counts are the YouTube API's own totalResults estimate, which Google documents as approximate, not an exact index count. They are intended only for rough comparisons between keywords and should not be quoted as absolute video counts.
• Videos are sampled by the API's own relevance ordering, not by view count, so demand describes the median inside the observed relevant sample rather than what the biggest channels achieve. The sample is the up to the first 50 results the API returns, not a random draw, so it inherits whatever ranking bias YouTube's search applies.

Method

Each keyword is searched twice against the YouTube Data API v3: once over videos published in the last 30 days, once over the block running from 90 to 60 days ago. The two windows never overlap. Statistics are requested for up to 50 videos the API returns first by relevance in each window, not the biggest by view count. Requested and returned detail counts are reconciled; partial or unavailable details cannot be ranked. Each sample is then split by recorded runtime: three minutes or less enters the short-form cohort, and over three minutes enters the long-form cohort. The split is done on the same data, not by a second search. Demand is sampled median video performance: median views-per-day of relevant recent videos within one length. Competition is estimated from the API's approximate totalResults, scaled by the relevant share observed in that format. Supply movement applies the same relevance rule to recent and comparison cohorts. Relative rank is calculated separately within each length. It orders today's eligible panel rows and is not an absolute score. For short-form (20 eligible rows), it is the leave-one-out residual of log(demand) against log(estimated on-topic competition), ranked within that length. For long-form (14 eligible rows), it is the leave-one-out residual of log(demand) against log(estimated on-topic competition), ranked within that length. Every positive verdict additionally requires configured absolute demand, sample, peer and channel-spread gates and must survive removal of its highest-performing video; Build uses the higher demand floor and top-quartile stability. Demand momentum compares today's sampled median against an accepted completed observation at the required time separation, matched by length and measurement population. This report's history spans 0 day(s) across 0 accepted completed observation(s). This run attempted 40 Search Queries call(s) from that separate bucket and 40 unit(s) from the general API bucket.

Demand is sampled median video performance: median views-per-day among relevant videos returned for this keyword in the last 30 days, split by length. Competition is estimated, not an exact count: the API estimate is scaled by the on-topic share observed in the same format. Supply movement applies that identical relevance rule to both recent and comparison cohorts. Relative rank orders the eligible keywords within its own format: a Short is scored against other Shorts, never against long form, because the two sit orders of magnitude apart on views per day. Rank is descriptive, not a mark out of 100; absolute demand, sample, peer, channel-spread and stability gates control the verdict. The verdict is a content-production call for today, not a recommendation to remove the keyword from the locked evidence panel. Keywords whose results were mostly off-topic are excluded rather than scored.

2 keyword/format row(s) could not be ranked. They are omitted from the table rather than shown with a guessed score.

• Caribbean slave rebellions history short-form: unranked: only 31% of results were on-topic, the search term is too ambiguous to measure (4 of 13 sampled videos were on-topic, 31%. Try a more specific search term.)
• Hattie McDaniel history long-form: unranked: only 33% of comparison results were on-topic, so supply movement is not like-for-like (Comparison window: 5 of 15 sampled videos were on-topic, 33%. Recent window: 9 of 20 on-topic, 45%.)

Highest observed videos in the relevant sample. Up to six per format are shown after collapsing near-identical titles and limiting any one channel to two rows. They are observations from YouTube's relevance-ranked sample, not proof of what universally "works". Every row links to its source and names the panel searches that found it.

Title words observed in relevant samples. This is a correlation, not a cause. Words and sampled median video performance are calculated separately for Shorts and long form. Nothing here shows that a word caused the views; the sample is relevance-ranked, not random. Treat these as format-specific prompts, not rules.

Written against the keywords that scored highest above. The evidence line under each is copied from the measured data, not written by the model.

Stated up front rather than implied by omission. Every item here is a real limitation of the data behind this report.

• Demand momentum is unavailable for seasonal conclusions: this niche requires 365 days of history and currently has 0. Once that horizon is reached, readings are compared at the configured 7-14 day separation. Until then the correct result is not enough data.
• 1 keyword/format pair(s) could not support a demand reading and are unranked: Caribbean slave rebellions history (short). The row-specific reasons distinguish too few relevant videos, low relevance, genuine zero, and incomplete evidence. A keyword can rank in one length and not the other; that is a real finding about where the topic lives, not a gap in the data.
• 2 keyword/format pair(s) have no comparable supply-momentum ratio because evidence was incomplete, relevance was insufficient, or the comparison denominator was a genuine zero: Caribbean slave rebellions history (short), Hattie McDaniel history (long).
• The short-form cohort is identified by recorded runtime of three minutes or less, because the Data API exposes no Shorts flag and no vertical/aspect field. Checked against YouTube's own classification on 150 real videos, that cutoff agreed 97.3% of the time: 4 wrong out of 150, none missed, and every error was a brief regular video assigned to that cohort. The check therefore observed a few percent of contamination in the short-form cohort and none in the long-form sample; it does not prove zero error outside that check. Videos whose runtime the API did not report, often live broadcasts, are counted in neither.
• Competing supply is reported per length, but the API only estimates a total for the keyword as a whole. The split between the two lengths is the share each took of the sample, applied to that estimate. It is an approximation laid on an approximation, and is the softest number in this report.
• Demand momentum compares the last-30-day cohort against the same cohort measured on an earlier run. Consecutive days overlap by 29 of 30 days, so day-over-day movement is heavily damped by design. Week-over-week change is the meaningful reading; a single day's shift is not.
• A keyword is only called 'build' if its on-topic long-form videos take a median of at least 20 views/day. For the short-form cohort the bar is 200 views/day instead, because the two lengths do not share a view scale and one floor for both would pass essentially every short-form row. A lower Test call still requires 5 long-form views/day and 50 short-form views/day. Today 13 of 19 short-form readings sat under it, and 20 of 20 long-form readings sat under it. Relative rank necessarily has a winner even in a uniformly weak set, so every positive call also needs enough absolute demand, relevant videos, peer keywords and channel spread, and must survive removal of its fastest video. These floors are judgement calls set in the niche config, not measured constants.
• Supply counts are the YouTube API's own totalResults estimate, which Google documents as approximate, not an exact index count. They are intended only for rough comparisons between keywords and should not be quoted as absolute video counts.
• Videos are sampled by the API's own relevance ordering, not by view count, so demand describes the median inside the observed relevant sample rather than what the biggest channels achieve. The sample is the up to the first 50 results the API returns, not a random draw, so it inherits whatever ranking bias YouTube's search applies.

Method

Each keyword is searched twice against the YouTube Data API v3: once over videos published in the last 30 days, once over the block running from 90 to 60 days ago. The two windows never overlap. Statistics are requested for up to 50 videos the API returns first by relevance in each window, not the biggest by view count. Requested and returned detail counts are reconciled; partial or unavailable details cannot be ranked. Each sample is then split by recorded runtime: three minutes or less enters the short-form cohort, and over three minutes enters the long-form cohort. The split is done on the same data, not by a second search. Demand is sampled median video performance: median views-per-day of relevant recent videos within one length. Competition is estimated from the API's approximate totalResults, scaled by the relevant share observed in that format. Supply movement applies the same relevance rule to recent and comparison cohorts. Relative rank is calculated separately within each length. It orders today's eligible panel rows and is not an absolute score. For short-form (19 eligible rows), it is the leave-one-out residual of log(demand) against log(estimated on-topic competition), ranked within that length. For long-form (19 eligible rows), it is the leave-one-out residual of log(demand) against log(estimated on-topic competition), ranked within that length. Every positive verdict additionally requires configured absolute demand, sample, peer and channel-spread gates and must survive removal of its highest-performing video; Build uses the higher demand floor and top-quartile stability. Demand momentum compares today's sampled median against an accepted completed observation at the required time separation, matched by length and measurement population. This report's history spans 0 day(s) across 0 accepted completed observation(s). This run attempted 40 Search Queries call(s) from that separate bucket and 40 unit(s) from the general API bucket.