Why source attribution beats black-box SEO scores

Most SEO tools hand you two numbers: a search volume and a difficulty score. Both are black boxes. You can't see where they came from, which sources they blend, or why the tool is confident. You're asked to trust a number with no provenance.

Here's what that costs, using two real searches from our Power Tools data.

Two searches. Same volume. Opposite realities.

Within 4% of each other on volume. Same niche. A volume-and-difficulty tool prints approximately the same row for both.

They are not the same opportunity. The first is a wall: Harbor Freight owns essentially all of the attention on a query about a product Harbor Freight sells. Ten channels are present and nine of them split 0.1% between them. The second is a live market — the leader holds 42%, and the rest is genuinely contested.

Make a video for the first and you are invisible. Make one for the second and you're in a fight you might win. The volume number is identical. The answer is opposite. You only find out which one you picked after you've made the video.

That's the whole argument, and everything below is about what it takes to be able to draw that table at all.

Keep the sources separate

The reason most tools can't show you that is that they blend first and report second. Once volume: 1900 / difficulty: 34 exists, the inputs are gone.

We keep three public signals apart and join them on the keyword, so every conclusion still carries what produced it:

• Search autocomplete — confirms people actually type the phrase, and tags what they want.
• YouTube — live attention in views per day, plus who holds it.
• Web results — which sites rank, and how concentrated they are.

Joined per keyword, not averaged into a score. The join is the product; the separation is what makes it auditable.

What a score would have swallowed: 585 rows

Concretely. YouTube view counts are a lifetime total, so demand has to be a rate — views per day since publish. That formula has a failure mode: a video with 2M views published nine days ago reads as 220,000 views a day, forever. It isn't a market signal, it's a viral spike mid-flight.

We zero those. The numbers:

Five hundred and eighty-five rows out of twelve million — five thousandths of one percent of the data — would have contributed 3.8% of all demand in the entire corpus. That's a ~760× over-representation, from rows you'd never find by eye.

A black-box score doesn't remove that. It doesn't tell you about it either. It averages it in and hands you a confident-looking integer.

Counting the same video eight times

A second, quieter one. Our YouTube data is keyword-grained: one row per (keyword, video). A good video ranks for hundreds of keywords, so it appears hundreds of times.

Across the corpus that's 12,371,436 ranking rows for 1,475,909 distinct videos — 8.38 rows per video. Sum demand naively and you inflate it more than eightfold, and you inflate it unevenly — the broadest-ranking creators get the biggest multiplier, which is exactly backwards from a fair comparison.

So every figure we publish is deduplicated per video first. It's an unglamorous correction that nobody would ever see in a score. It's also the difference between a leaderboard and a fiction.

When two sources disagree, the disagreement is the signal

This is the part that a blended number cannot do at all.

Two independent questions decide whether a niche's traffic is commercially worth anything. Does the audience arrive wanting to buy — measured from keyword intent? And are creators wired to capture it — measured from affiliate and commerce links on the videos holding attention?

Blend them and you get one lukewarm score. Keep them apart:

Look at Power Tools. By keyword intent it's an 18% — barely commercial, skip it. That reading is wrong, and we know it's wrong because the other source contradicts it: creators affiliate-link 53% of the attention in that niche, and they don't do that for fun.

The explanation is that tool queries are model numbers — DeWalt DCD777 — with none of the best / vs / review words a keyword-intent regex looks for. The intent signal has a blind spot, and the wiring signal sees straight through it.

A single blended score would have averaged 18% and 53% into a meaningless middle and told you nothing. Two attributed sources told you which one to distrust, and why. You cannot get that from a number that has forgotten where it came from.

It also catches our own mistakes

The honest version of this argument is that attribution isn't just for auditing other people's numbers. It's how you catch your own.

Two from the last month:

We published a wrong number and the evidence caught it. Our attention model multiplies views by video length. Uncorrected, that credits a three-hour podcast with three full viewer-hours per click, and it made the >2h band look like 34% of all watch-time. Applying a completion discount halved it to 17%. The error was visible because the inputs were still separable — views here, duration there, discount explicit. Baked into a score, it would have been unfalsifiable. (The full write-up.)

Our own data has junk in it, and we can see it. Searching our Power Tools data for high-demand, single-channel keywords surfaces jigsaw explorer auto solver — where the top channel is Mark Rober — and jigsaw y, won by the K-pop group IVE. Jigsaw is a saw and a puzzle; impact is a driver and a noun. Real polysemy, real contamination, sitting in a real production table.

We'd rather tell you that than not. A tool that reports a clean-looking score for jigsaw y isn't more accurate than us — it's just not showing you the part where it was wrong. We run an exclusion list and an AI relevance pass over the head of the queue, both imperfect, and the numbers in this post are filtered through them.

What this doesn't get you

Attribution isn't free and it isn't magic.

It's more to read. A score is one integer; a source-attributed row is three signals, a concentration figure and a caveat. That's more work, and for a throwaway question it's more work than the question deserves.

It's not clairvoyance. We can tell you a search is contested and by whom. We can't tell you your video will win it.

And some of it is modelled, not measured — the completion curve above is calibrated against published studies, not observed, because YouTube doesn't publish retention. Where that's true, we say so on the number itself rather than in a footnote nobody reads.

The point

The goal was never a prettier dashboard. It's that every claim in your report can be walked back to a source you can check — and that when two sources disagree, you get to see the disagreement instead of receiving its average.

volume: 1900 / difficulty: 34 can't tell you Harbor Freight owns 99.9% of a query. It can't tell you 585 rows were about to skew your dataset. It can't tell you which of its own inputs is blind. It was never going to, because it threw all of that away before it reached you.

Every figure here is from our production data and traces back to specific videos and searches. Poke at it yourself — the niches we cover, or check a channel. No signup.