Where keyword tools get their numbers

I run keyword research across a portfolio of sites, and for years I treated the number in the volume column like it was a fact. It isn’t. It’s an estimate built from a patchwork of data sources, and once you know what’s underneath it, you stop making decisions you’ll regret.

The honest starting point: nobody outside Google has the real numbers

Google doesn’t publish exact search counts for the open web. Every third-party tool, and even Google’s own free tools, are reconstructing volume from partial signals. Some of those signals are closer to the truth than others. None of them are the truth.

This matters because I’ve killed content plans based on a “2,400/mo” search volume that turned out to be almost entirely branded traffic for a completely different query, bucketed into the same number by whatever tool I was using. The number wasn’t wrong exactly. It was measuring something I didn’t understand.

Google Keyword Planner: real data, deliberately blurred

Keyword Planner pulls from actual Google Ads auction data: real searches tied to real ad impressions and clicks. That’s the closest thing to ground truth you’ll find in a free tool.

The catch is that Google rounds and buckets these numbers on purpose, especially for accounts that aren’t spending real ad money. Instead of “1,847 searches,” you get a range like “1K-10K.” That range is wide enough to be nearly useless for comparing two similar keywords. I’ve seen a term that clearly gets more traffic than another sit in the same bucket, because Keyword Planner isn’t built for content research, it’s built to help advertisers set bids.

If you spend meaningfully in Google Ads, the ranges tighten up and you start seeing closer-to-exact figures. That’s not a bug, it’s the business model. Google wants advertisers to have clean data and everyone else to have a rough sketch.

Keyword Planner also averages volume over the trailing 12 months by default, which flattens seasonality. A term that spikes hard in December and dies the rest of the year shows up as a flat, unremarkable monthly average unless you dig into the monthly breakdown.

Clickstream data: the backbone of most paid tools

Ahrefs, Semrush, Moz, and similar tools rely heavily on clickstream data: anonymized records of what people actually typed and clicked on, collected from browser extensions, toolbars, and data partnerships. A panel of users installs something, that something logs their browsing and search behavior, and the vendor buys or licenses that log in bulk.

This is why volume numbers differ between tools that are all trying to measure the same thing. Ahrefs’ panel isn’t Semrush’s panel. Different panel size, different geographic mix, different browser and device skew. If your keyword tool’s panel is heavy on desktop Chrome users in the US, and your actual audience is mobile Safari users in Southeast Asia, the model is extrapolating from a sample that doesn’t look like your market. I’ve watched volume estimates for the same term in a non-US country vary by a factor of three between tools, purely because of panel composition.

Clickstream data is also inherently a sample, not a census. Nobody’s clickstream panel captures more than a small slice of total search traffic. The rest gets modeled statistically: known clickstream volume for term A, known relationships between term A and similar terms, extrapolate to term B. That modeling step is where a lot of the disagreement between tools comes from. It’s not that one vendor has better data and another has worse data across the board. They’re running different statistical models on different partial samples.

Search Console: the only numbers that are actually about your site

Google Search Console shows real impression and click data, but only for queries where your pages already show up in search results. It’s not a discovery tool for new keywords, it’s a verification tool for keywords you’re already ranking for.

I treat GSC as the tiebreaker. If a third-party tool says a keyword gets solid volume but GSC shows near-zero impressions for a page that ranks reasonably well for it, I trust GSC. It’s not sampled or modeled, it’s what actually happened on my own site. The tradeoff is that GSC can’t tell you anything about a keyword you don’t rank for yet, so it’s useless for finding new opportunities. Its value is in checking the third-party estimate against reality once you’re already visible.

Autocomplete and “People also ask”: ideas, not numbers

Autocomplete suggestions and the People Also Ask box are scraped by a lot of keyword idea tools to generate lists of related queries. This is a legitimate way to find phrasing you wouldn’t have thought of, because it reflects what people actually typed into the box, not a model’s guess at what they might type.

But there’s no volume attached to any of it natively. Autocomplete surfaces a term because enough people type it, not a specific count of how many. Any volume number sitting next to a scraped autocomplete suggestion in a tool’s UI is still coming from the clickstream or Ads-data pipeline described above, bolted onto a list generated by a completely different method. Worth knowing so you don’t assume the whole tool runs on one clean data source.

Why the same keyword shows different numbers in every tool

Put the pieces together and the disagreement makes sense. One tool leans more on Ads auction data, another leans more on its own clickstream panel, a third blends multiple data licenses and reweights them with its own model. Update frequency differs too, some tools refresh volume monthly, others quarterly, and a term that’s trending right now might still show last quarter’s number in a tool that hasn’t refreshed yet.

None of this means the tools are useless. It means the absolute number is the least reliable part of the output. What tends to hold up better across tools is the relative comparison: keyword A consistently shows more volume than keyword B, even if the exact figures differ. I plan content around that relative ordering and around trend direction, not around the specific number in the column.

How I actually use volume data day to day

For new content, I use a paid tool’s volume as a rough sizing signal, mainly to avoid writing 2,000 words for a term nobody searches for. I don’t treat the number as precise enough to choose between two similarly-sized keywords based on a small gap.

For anything I already rank for, GSC impressions are the real signal. If GSC shows a page pulling steady impressions for a query, that tells me more about actual demand than any third-party estimate, because it’s not a sample or a model, it’s what happened.

For seasonal or trending terms, I pull the monthly trend line instead of the averaged annual number, in whichever tool shows it. The averaged number hides exactly the pattern I need to see if I’m trying to time content around a seasonal spike.

I don’t chase precision on volume the way I used to. The number is a directional estimate built from partial data and statistical modeling, not a headcount. Once that’s the working assumption, the tools become more useful, not less, because you stop trusting the decimal point and start trusting the shape of the data.

If you want more of this kind of breakdown on how the tools you already use actually work under the hood, head back to the homepage for the rest of what we’ve published.