The number next to a keyword in any research tool looks like a measurement. It is a model output, and knowing how it is built changes how you use it.
The upstream source
Nearly every commercial keyword tool traces back, directly or indirectly, to Google Keyword Planner, supplemented by clickstream panels and the tool's own modeling. Keyword Planner was built for advertisers, and its numbers carry three properties that matter for SEO:
It is bucketed. Keyword Planner reports volume in ranges, and accounts with low spend see coarser buckets than accounts with high spend. Tools that resell this data smooth the buckets into precise-looking integers. A keyword reported as "1,300 searches" may sit anywhere inside a range that the tool has picked a midpoint from.
It is a twelve-month average. A term that gets 12,000 searches in one month and near zero for the rest of the year reports as roughly 1,000 a month, every month. For anything seasonal — tax, insurance renewals, holiday retail, weather-driven trades — the annual average describes no month in the actual year.
It aggregates near-variants. Close variants are grouped, which means the reported number covers a family of phrasings rather than the exact string you typed. This inflates apparent volume for the specific phrase you are targeting.
The clickstream layer
Tools supplement Planner data with clickstream panels — anonymized browsing data from browser extensions, apps, and ISP-level partnerships — then extrapolate from panel to population. Extrapolation from a non-random panel is where most of the error enters. Panels skew toward certain demographics, devices, and regions, and the correction factors are proprietary.
This is why two reputable tools disagree by a factor of two or more on the same keyword. Neither is lying. They are running different models over partly different inputs.
What the error means in practice
For SEO decisions, the practical consequences are narrower than the error bars suggest, because most keyword decisions are comparative rather than absolute.
Comparative use is fine. If tool A says keyword X gets 4,000 searches and keyword Y gets 200, the ordering is almost certainly right even if both numbers are wrong. Prioritizing X over Y is a sound decision. Rank-ordering within one tool is the single most reliable thing keyword volume data does.
Absolute forecasting is not fine. Multiplying a modeled volume by a borrowed CTR to promise a traffic number produces a figure with compounding error in a single optimistic direction. If you must forecast, forecast a range, state the assumptions, and treat the output as a scenario rather than a target.
Zero-volume keywords are not zero. Tools report zero or "—" when a term falls below their reporting threshold, not when nobody searches it. Long-tail terms with real, high-intent demand routinely show as zero. For local and highly specific B2B terms, the zero-volume bucket is frequently where the actual conversions live.
Better signals when volume fails
When volume data is thin or you suspect it, there are more direct signals available:
- Search Console impressions. If you rank at all for a term, impressions are a floor on real demand — and they are counted, not modeled.
- Google Trends. Relative and directional rather than absolute, but built on actual query logs, which makes it the right tool for seasonality and for the question "is this growing."
- Autocomplete and People Also Ask. Presence is evidence that a query is common enough for Google to have surfaced it. No number attached, but a real signal.
- Your own site search and support tickets. The highest-intent phrasing your customers use, in their own words, with zero modeling in between. Chronically underused.
The reporting rule that keeps you honest
Report volume with the tool name attached. Not "this keyword gets 2,400 searches a month" but "Semrush estimates roughly 2,400 US searches a month for this keyword." The attribution does two useful things: it signals that the number is an estimate from a specific model, and it makes the inevitable disagreement with a client's own tool a conversation about methodology rather than an accusation of error.
The teams that get the most out of keyword data are the ones that treat it as a prioritization input rather than a promise. Rank the opportunities, pick the ones where the ordering is clear, and let the actual impression data from Search Console correct the model as it arrives.