Third-party traffic figures are built by taking a keyword corpus, estimating each keyword’s search volume, working out roughly where a site ranks for each, and applying an assumed click-through curve. Four layers of estimation, compounding.
So the honest answer to “how accurate is it” is: directionally good, individually unreliable, and systematically biased in ways that are actually quite predictable. That last part is what people miss, and it is the part that makes the data worth paying for.
How the number is actually built
It helps to know exactly where the error enters, because then you can predict which sites it will hit hardest. Every estimate on the market is assembled the same way:
- A keyword corpus. A fixed list of queries the tool tracks — tens or hundreds of millions of them. That sounds exhaustive until you remember that a large share of what Google handles each day are queries it has never seen before.
- A volume model for each keyword. Usually clickstream panels, calibrated against whatever ad-platform figures the vendor can obtain. Panels skew by country, device and demographic, and that skew is inherited wholesale.
- A rank position for the site, per keyword. Crawled on a schedule. It is a snapshot taken every few days or weeks, not a continuous recording, and Google is not serving one fixed result set to begin with.
- An assumed click-through curve. A share of clicks by position, applied more or less uniformly across every query type.
Multiply those together, sum across the corpus, publish. Four models stacked, each carrying its own error, and the product inherits all of them at once.
Where estimates go wrong
- Branded traffic is undercounted. Models are built on non-branded keyword corpora, so a site with a strong brand looks smaller than it is. The stronger the brand, the worse the gap.
- Long-tail is undercounted. Queries too rare to appear in the corpus are invisible, which penalises exactly the programmatic and Q&A sites that live on them.
- Non-search traffic is absent entirely. Newsletter, social and direct do not appear. A site with a big audience and modest SEO reads as small.
- Small numbers are noisy. Below a few thousand visits, a single ranking change swings the estimate wildly.
- SERP features break the click curve. An AI overview, a featured snippet or a full-width pack changes the real click share for that query dramatically. The assumed curve does not know it is there, so position three on a feature-heavy SERP is scored as though it were position three on a plain one.
Notice that four of these five push the same direction: estimates tend to undercount rather than overcount. That systematic bias is what makes them useful.
Consistently wrong in the same direction is far more useful than occasionally wrong in random ones.
Which way the error runs, by site type
Because the causes are known, the direction of the error is predictable. This is the mental adjustment worth making before you read any row in any tool:
| Site type | Likely direction | Why |
|---|---|---|
| Brand-led ecommerce | Undercount, large | Most of its demand is branded and outside the corpus |
| Programmatic / database | Undercount, large | Lives on tail queries the corpus never sampled |
| Forum or UGC | Undercount, large | Same tail problem, plus heavy direct traffic |
| Affiliate in a mapped niche | Close to accurate | Head terms are well covered and well modelled |
| News and trending | Unstable either way | Rank snapshots miss the spikes that carry the traffic |
| Local, multi-location | Undercount, moderate | Map pack behaviour is not in the click curve |
A DR 14 site showing 80,000 visits is therefore most interesting when it sits in one of the undercount rows, because the real figure is probably higher still.
Decisions that survive the error
Anything comparative. If the same model is applied to every site in the index, then site A reading twice the size of site B is meaningful even when both absolute numbers are off. The error largely cancels.
That covers most of what this data is actually for: ranking sites within a niche, spotting outliers against their authority, measuring before-and-after around an update, and comparing how two niches behave. All comparative, all robust.
It also covers the shape of a trend. A site that has tripled over eighteen months has tripled, whatever the baseline was, because the same four models produced both ends of the line. Shape survives even when scale does not.
Decisions that do not survive it
Anything where the absolute number is the decision. Do not value an acquisition off an estimate. Do not forecast ad revenue from one. Do not tell a client they will get exactly 40,000 visits. For those you need the seller’s first-party analytics — a Search Console export beats any model — and if they will not show you, that is information too.
The dangerous cases are the ones where an estimate gets laundered into a spreadsheet and quietly loses its error bar. A modelled figure multiplied by an RPM produces a revenue number that looks precise to two decimal places and is not precise at all.
A sixty-second sanity check
Before you act on any single row, run this. It costs a minute and catches most of the bad reads:
- Look at the trend line before the number. A smooth eighteen-month climb is a real site; a vertical cliff is usually a tracking artefact.
- Check the ratio of ranking keywords to estimated traffic. Thousands of keywords and very little traffic means tail-heavy, which means undercounted.
- Search the brand name. Real branded volume tells you the model is missing a chunk.
- Open the top pages. If they are not the pages you would expect to carry the site, the rank data is stale.
- Pull the same domain in a second tool. The gap between them is your practical confidence interval.
When two tools disagree
They always disagree, and the disagreement is informative rather than embarrassing. Two tools differing by 20% are working from different corpora and broadly agreeing. Two tools differing by 5x are telling you the site’s traffic comes from somewhere neither model covers well — which is itself the finding. Sites that break the models are frequently the interesting ones.
The mistake is picking whichever figure supports the decision you had already made. If you are going to average them, decide that before you look.
Our own rules
Two choices we made for this reason. First, the index has a floor of 1,000 monthly visits — below that the noise exceeds the signal and the row would be misleading rather than merely imprecise. Second, we show update impact as a percentage change rather than an absolute delta, because the percentage is the part that holds up.
We also say all of this openly on the FAQ and the update methodology, which we would rather do than let someone discover it during a deal.
The practical takeaway
Treat every traffic figure as a rank, not a measurement. Ask “is this bigger or smaller than that, and by roughly how much” and the data is excellent. Ask “exactly how many people visited” and no third-party tool on the market can honestly answer you.
Frequently asked questions
How accurate are third-party traffic estimates?
Directionally good and individually unreliable. They are built from four stacked models — keyword corpus, search volume, rank position and an assumed click curve — so each figure carries all four error bars at once.
Why do two SEO tools show different traffic for the same site?
Because they use different keyword corpora and click curves. A 20% gap means they broadly agree; a 5x gap usually means the site earns traffic from somewhere neither model covers well, such as branded search or a very long tail.
When should I not rely on a traffic estimate?
Whenever the absolute number is the decision. Do not value an acquisition, forecast ad revenue or promise a client a specific visitor count from a modelled figure. Use estimates to rank and compare, never to measure.
The takeaway An estimate is a position in a league table wearing the costume of a measurement. Use it to sort, compare and rank — never as the input to a number someone will sign against.