Where our search numbers come from, and what they don't tell you

How do we produce a search volume?

We start from a service in our catalogue — "botox", "dermal fillers" — and expand it into every related phrase a commercial search-data provider knows about. For Botox that returned 320 phrases. Each comes back with a modelled monthly volume for a geography: in our case the metro our catalogue resolves for GTA markets.

The word modelled is doing real work there. Nobody counts every search. Providers infer volume from clickstream panels, ad-platform ranges and their own smoothing, then publish a single number that looks far more precise than the method behind it. When we print 5,400, the honest reading is "a few thousand, and more than the phrase below it".

That is why every estimated figure on this site renders with a striped bar. It is the same encoding our dashboard uses, and it means the same thing in both places: this is modelled, not counted.

Why do most phrases come back as zero?

Of 320 Botox phrases, 262 have no recorded monthly volume at all. Providers suppress numbers below a reliability threshold rather than publish noise, so zero means "too rare to measure confidently", not "nobody searches this".

For classic search that was a rounding error. For AI answers it is the whole tail. People ask assistants in full sentences — "is masseter Botox safe if I grind my teeth at night" — and a full sentence essentially never matches a measured keyword string.

So we use volume to rank concerns, not to size markets. A phrase with zero volume still tells us a worry exists; it just cannot tell us how common it is.

What do we throw away, and why?

Grouping phrases into concerns is a judgement, so it gets reviewed by hand and the review is logged. On the most recent pass we kept 68 groupings and merged 15 away as near-duplicates — "dark circles" folded into under-eye, "thin lips" into lips, "fix bad filler" into dissolving.

Merging matters because an ungrouped list flatters itself. Split one concern across three names and each looks small; merge them and the real shape appears. Every merge is recorded with its reason so the count can be argued with rather than taken on trust.

What can these numbers not tell you?

Four things, and we would rather say them here than be asked. They cannot tell you what any individual assistant is asked, because no provider measures assistant traffic and we do not pretend to. They cannot tell you who gets named in an answer — that is a different measurement, and it is what Radar exists to collect.

They cannot be read at neighbourhood level: keyword providers target metros, so a figure we attach to a GTA market is measured across a wider area than the market itself. And they are a snapshot, not a trend: we re-pull each quarter, and a number that moves between reports may be the market moving or the provider's model moving.

Where a limit like that changes what a report means, we put it in the report, not in the footnotes.

Why publish the method at all?

Because we are asking small business owners to act on numbers they cannot independently check, in a field where a lot of people are selling certainty they do not have. The only honest answer to that is to show the working and mark the estimates.

It is also the standard we hold client sites to. We tell every client that a statistic without a source line is not worth publishing. It would be strange to exempt ourselves.

Common questions