YouTube Keyword Research

Give it a topic in plain words. It expands that into candidate keywords, runs a live query for each against the 687M videos we index, drops the ones whose matches are off topic, and returns a filter plus the channels or videos it selects.

What you get

• A filter set, always. Reusable keyword groups, a link that opens a report with the filter applied, and the Boolean expression so anyone can re-run or audit it.
• Trend data. Matching uploads, sorted and windowed, with prevalence numbers: how many videos, how many distinct channels.
• Channel targets. Channels tiered by how much they cover the topic, ranked, context-validated, and flagged for whether you can book them.

Why a synonym list falls short

What people use it for

A brand sizing a niche

> how big is the home espresso topic on YouTube, and is it growing?

Prevalence numbers and the matching uploads by date.

An agency building a target list

> find channels for a retirement client, deep, channel targets

Tiered, context-validated channels with subscribers, price, and outreach email where on file.

A creator-sourcing analyst

> keywords that represent budget meal prep content

A filter set, and the report link showing where the list came from.

Someone chasing a launch

> make a topic for this week's product launch, run autonomously

A web lookup for the real product names, then expansion covering its siblings.

A researcher writing it up

> how common is each of these keywords?

Counts per keyword, ranked, with the scope stated.

How it works

Seven stages. Scripts do the deterministic work at no model cost; classification runs on small fast models in batches.

1. 0

   Stage 0: Set up

   Your sentence is kept verbatim as the intent, since every later check is judged against it. Scope, operator and a breadth judgment are stated back before anything is spent.

   One cheap probe asks whether a channel-identity version of the topic exists, or whether the market is channels that keep returning to it.

   runs expand_entities.py

2. 1

   Stage 1: Expand

   Each named entity expands into its family: company, product line, version, codename, and the sibling that shipped alongside. 8 to 15 candidates for a narrow niche, 30 to 50 for a multi-facet event.

   A shallow synonym list is the biggest failure mode. Coverage is won or lost here.

3. 2

   Stage 2: Probe

   One live query per candidate returns matching documents, distinct channels reached, and real samples. Recency rides the same query free, flagging terms that live off an old back catalogue.

   Samples collapse to distinct channels, so one prolific creator cannot flood the evidence.

   runs probe.py

4. 3

   Stage 3: Validate

   Each keyword's samples are checked against your intent, and off-topic terms are dropped with the reason shown. At scale a batch classifier does it, with a completeness check that re-sends anything left out.

   Relevance is judged from samples. Scope is your call, so those come back with snippets attached.

   runs select_keywords.py

5. 4

   Stage 4: Refine

   At least three rounds of recomposing the filter. Each measures coverage, scores fitness, records the query, and names its move: narrow, expand, or backtrack.

   A move that reduces fitness is discarded and a different axis tried.

6. 5

   Stage 5: Materialize

   One aggregation tiers every matching channel by how much of its output is the topic. Results are then built from the final filter verbatim.

   The tier table is 2 to 3 queries however many channels match, so it runs before any per-channel spend.

   runs search_channels.py --intensity

7. 6

   Stage 6: Deliver

   Filter set, report link, results, and the Boolean expression. Anything pruned or translated is listed. Nothing is saved unless you say so.

   runs build_report.py

Every spelling is a different word

The index lowercases and splits on whitespace and punctuation, but does not split letter-digit runs and does not stem. One product name written four ways produces four sets of tokens.

Four populations, one product. All-time counts, verified live in our index when the skill was written, so they drift as new videos land.

So anything with a number, version or model name is probed in every spelling, and the survivors fold into one group. Inflections too, since nothing is stemmed: invest, investing and investments are three terms.

When the model does not know the names

For a topic newer than the model’s knowledge, recently renamed, or thick with insider jargon, a web lookup runs first in an isolated context. It returns names only. Each then earns its place through the same probe and validation as any other candidate, and a hallucinated one probes to zero and drops.

The word is there. Is the topic?

Every probe returns two counts: total documents, and distinct channels. Judge a niche by the second, since one prolific creator can produce a thousand matches alone.

Samples are then read against your sentence. Two verified examples from one retirement run:

• "social security" reaches 228 channels, a good share of them Brazilian pension-law firms. Excluding INSS leaves 192.
• annuity reaches 596, including finance coursework: annuity due, ordinary annuity, present value. Excluding that cluster leaves 557.

Neither term is wrong, and neither problem is visible until you read what the term matched.

Whether matches are on topic is judged from evidence. Whether a sub-topic belongs in your scope is your call, so those come back with snippets and an in-or-out question.

Content fields

Ranking weights them title^4, summary^2, transcript^1. A term too noisy corpus-wide is often precise in titles alone, so rather than drop it, the skill probes per field and keeps the group with a field restriction attached. Fields are set per group, not once for the filter.

Boolean queries a list cannot express

Every group is a self-contained Boolean query. Groups combine with OR; inside one you get required terms, alternatives and exclusions with real precedence.

Scope an exclusion to its own arm

A whole-filter exclusion hits groups that never had the problem. Inside its own group, it de-noises only the family that needed it:

("mythos 5" | mythos5) -keto

On one live run this kept 51 on-topic documents a global exclusion had removed.

Guard against over-exclusion

Re-run the core with and without the exclusion; the count should barely move. Excluding bare film from a Cannes Lions core cut it 24%, since on-topic content shares the word. The phrase "film festival" cut 4%, so that is the one that ships.

Anchor a root too broad to use alone

A place name has no single polluter to cut. Require an anchor token plus a domain qualifier:

cannes +lions +(advertising | agency | campaign | "young lions") -"film festival"

"cannes lions" fires only when the words sit side by side; cannes +lions fires when both appear anywhere, catching “Young Lions at Cannes”. The qualifier supplies the precision, which bare cannes +lions lacks. Measured live, that group added roughly 815 distinct channels beyond the phrase, around 80% on topic.

Judge a broad term on its residual

Probe the candidate with the core phrase excluded and read those samples. Relevance scoring otherwise floats covered documents to the top and hides how little is new.

Research operators never ship

* catches inflections, ~1 absorbs typos, "a b"~2 catches near-phrases. None survive into a report filter, so variants are enumerated first: retire* becomes (retire OR retiring OR retirement).

The grammar translation

A report link uses the uppercase words AND, OR and NOT; raw |, + and - are literal text there. Groups are translated on the way out:

("fable 5" | fable5) -keto

("fable 5" OR fable5) AND NOT keto

Every translation is listed, and an unsupported operator fails loudly. This is the most common way a hand-built filter breaks: the link looks right and is matching a literal plus sign.

Channels tiered by coverage

Few channels are entirely about any topic, so “channels about X” alone returns a near-empty list for most niches.

The top tiers are context-validated: the text around each keyword occurrence is classified on topic, mixed or off topic, with an evidence quote and confidence. Mixed channels are kept and labelled, only clear misses excluded, and the excluded list shown. Channels you cannot book are flagged, not filtered out.

If a topic has almost no core channels, you are told, with the recurring tier framed as the market. A near-empty list is never padded.

Run modes

Say which you want, or answer when it asks; it will not pick silently. Say run autonomously to skip every check-in, which implies deep with both result types. Spend is narrated at every stage transition. Report links are free, and so is asking the skill to explain itself.

Try it

uv tool install thoughtleaders-cli
tl auth login
tl setup claude

The skill arrives as a slash command. The full setup page covers other installers, headless machines, and the one-liners for Codex, Gemini CLI and Cursor.

Start on a topic you know well, so you can check the output against your own knowledge.

Good next prompts:

• how does tl-keyword-research work? for the built-in guide, free
• keywords that represent budget meal prep content, quick for a starter filter
• how big is the GRWM trend on youtube right now? trend data only for prevalence, no channel list
• find channels covering Medicare for a retirement client, deep, channel targets
• make a topic for cannes lions, run autonomously, both for no pauses

Options, in plain words

Scope and limits

YouTube uploads only, longform by default. Shorts and live on request, and the scope in force is echoed back with every result.

Headline counts are all-time. Recency rides the same query, so a term living only in an old back catalogue is tagged stale.

There is no universal right result size. A few dozen channels is complete for a narrow niche and a miss for a broad one, so breadth is stated up front and measured at the end. If mining runs dry, a small result is reported as correct, not padded.

The corpus is the boundary. This searches what we index, titles, creator descriptions and transcripts across 1,302,318 tracked channels, not the open web. It is the same data our sponsorship matching runs on.