SourceSignal Research · Search and YouTube market data, separate from the product reports
Here for product reviews? How product-review analysis worksHow these numbers are made
Panel figures below are a snapshot taken on . The product reports are a separate dataset, with their method on each published example.
What we measure
The corpus behind our market data (the market reports, tools and charts) as of 10 September 2026.
Product niches tracked end to end, from Home Fitness to Fishing Gear.
Entities within those verticals: the grain most analysis actually runs at.
Intent-tagged search phrases expanded from seed terms through public autocomplete.
Observations of which video held which position for which keyword.
Distinct videos seen ranking, each enriched with its own view and duration data.
Distinct creators observed ranking for at least one tracked keyword.
Repeat measurements of the same videos over time: the basis of every movement figure.
These are counts of observations we hold, not estimates or projections. Collection changes over time; the date above is the version this page describes.
Where the signals come from
- 1
Seeds expand into keywords
Each sub-niche starts from a small set of seed terms, which we expand through public search autocomplete into the phrases people are actually offered. Every resulting keyword is tagged by intent: a “X vs Y” comparison is a different kind of search from “what is X”, and they are never pooled.
- 2
Results are collected per keyword
For each keyword we record which videos rank on YouTube and in what order. Each result is stored as a timestamped observation attributed to the keyword and source it came from, so any figure can be traced back to its evidence.
- 3
Videos are enriched
Ranking videos are measured individually for view count, publication date and duration: the inputs to demand, and the inputs to duration-weighted proxies rather than observed watch time.
- 4
The panel is re-measured on a scheduled cycle
Eligible queries are scheduled at 7-, 14- or 30-day intervals by collection tier. New snapshots support weekly comparisons when matched coverage is sufficient. A failed refresh can leave the last available week in place; the displayed window is the reference for freshness.
The formulas applied to this data (demand, concentration, authority, opportunity) are written out below →
The math behind the numbers
No black-box scores. Every headline metric is a formula you can read: here’s exactly how each one is computed, and why.
Noise-corrected lifetime view average
The demand field is lifetime views divided by video age at collection, with configured spike exclusions. It is a historical average, not views gained today or views attributable to search. Deduplicated aggregates count each video once within their stated scope.
How concentrated the tracked field is
The Herfindahl index of creator demand-shares: 0 = fragmented, 1 = monopoly. Its inverse N is the equivalent number of equally sized channels. Fragmentation does not establish an opening for a new entrant.
Demand vs. competition
This heuristic ranks the historical view-average signal per observed competing channel. It does not measure unmet demand, content quality, ranking difficulty or the chance of entering successfully.
Query wording used for collection priority
These are collection-priority weights assigned from query wording. A comparison receives five times the informational weight; that is a scheduling choice, not measured purchase intent or commercial value.
Comparable across niches
Markets span orders of magnitude, so counts are log-scaled to a 0–100 shape and demand bands come from global quantiles — every niche measured on the same ruler.
These are the same formulas behind every market report and the free tools; nothing is hidden or proprietary-by-obscurity. They run on the panel above, within the limits set out below.
The attention model
We store video duration and view counts, which allow us to construct duration-weighted attention proxies. These models use assumed completion; we do not observe viewers’ watched hours. They aren’t in reports yet; we’re showing the working while we build it.
Duration-weighted view proxy
Multiplying the historical view average by full video duration produces a duration-weighted proxy. It assumes every view spans the full video and is not actual watched hours. Deduplicate within the stated panel.
Assumed completion by video length
This proposed model applies an assumed completion fraction using length L in minutes. The constants are modelling choices, not observed retention for these videos. Actual watch time requires audience-retention data that we do not hold.
Who counts as a competitor
This proposed gate selects channels by observed video count, keyword footprint and category focus. Query association is not verification that every video reviews a product, nor proof that a channel is a suitable partner.
Aggregate modelled attention proxy
This sum combines the duration-weighted historical view average with assumed completion. It is a modelled panel proxy, not measured watch time, economic size, revenue or consumer demand.
The exploratory analysis of duration bands and their assumptions is in the write-up →
Rules we hold ourselves to
Most of the ways this data could mislead come from a handful of specific mistakes. These are the rules that prevent them, applied in the pipeline, not just in the writing.
We never report search volume
We do not hold query counts. The demand field is lifetime video views divided by age at collection, with spike exclusions. This historical view average includes views from any source; it does not measure current daily traffic, search-attributed views or buyer demand.
Search results include more than the niche
A niche's view figures count every video in the top ten results of our searches for that niche. YouTube mixes unrelated videos into those results, often viral Shorts, so the figures describe what our searches observe, not the niche's own audience. A check on one niche in October 2026 found a large share of the counted views were on videos not about it. A relevance filter is being tested; it is not applied to published figures yet.
Movement is always relative to its market
Movement figures compare matched-video changes with a same-week tracked-market benchmark. This reduces a common scan-window effect; it does not eliminate all sampling or measurement bias. Category indices chain relative ratios over their displayed window; creator momentum is a separate one-week comparison.
Matched pairs only
A video counts toward a movement figure only if we measured it in both weeks being compared. Comparing totals instead would measure how much we happened to scan that week, not how the market moved. This is why our movement numbers are ratios over a stable cohort rather than sums.
Complete weeks only
The current calendar week is withheld. Publication also requires the panel minimum for the relevant comparison. A complete calendar week does not guarantee complete collection; use the latest available week shown on each page.
Each video counts once
A video ranking for eighty keywords is one video, not eighty. Counting per keyword-appearance would inflate demand roughly eightfold: a large and very easy mistake to make with this shape of data.
Phantom spikes are zeroed
The demand calculation zeros videos above one million views when their publication data is missing or they were under 30 days old at collection. This is a noise-control rule; it does not prove that those views were artificial.
What these numbers cannot tell you
We see 15 niches, not all of YouTube
Every figure is scoped to the verticals we track. A creator’s standing is their standing within a measured market, not a platform-wide ranking. Channels outside our seeded keyword space are simply not observed.
Brands are extracted, not curated
Brand names are currently pattern-matched out of keywords and channel names rather than maintained as a verified registry. Brand attribution can be incomplete or incorrect and should not be treated as verified, and we do not publish them as headline numbers.
The history is young
Weekly view snapshots began accumulating recently, so trend figures rest on a handful of clean weeks. They are enough to rank movement within a week; they are not enough for seasonality, and we do not present them as such.
We measure reach, not revenue
Views, ranking positions and video durations are observable. Actual watch time, conversions, revenue and audience intent behind a view are not. We do not model them, and any figure implying we do is being read too generously.
Engagement is partial
We collect view counts and durations, not likes or comments, so we cannot compute engagement-rate metrics that depend on them.
Using these figures
Our figures are free to quote, in research, reporting or your own writing, with attribution and a link. We ask two things, both of which follow from the rules above: quote the figure with the date it was taken, and describe movement as relative to its market rather than as absolute growth.
SourceSignal, Methodology, 10 September 2026. Based on a panel of 79.0M YouTube ranking observations across 15 product niches.
Questions about the method?
If you are citing this work and need detail we haven’t published (window definitions, per-niche panel sizes, how a specific figure was derived), ask and we’ll answer.
About SourceSignal