Semrush Traffic Analytics, and what the numbers are actually worth

Traffic Analytics will tell you how much traffic any website gets. It will be wrong, and Semrush is more honest about that than the people quoting it.

In my own walkthrough the figure comes up on screen and the narration says it plainly: this is an estimated guess, it can be a bit more or a bit less. That sentence is the whole story of this tool, and it disappears the moment somebody screenshots the number into a pitch deck.

What it is actually measuring

Nothing, directly. It has no access to a competitor’s analytics, so it models visits from panel data, clickstream providers and its own crawl, then presents the output at the same visual confidence as a measured figure.

What you get per domain: total visits, unique visitors, pages per visit, average duration, bounce rate, desktop against mobile, traffic by channel, and the countries the traffic comes from.

The channel split is the most useful part and the least screenshotted. It’s where you find out that a competitor you assumed was winning on SEO is actually buying most of their traffic. In that same recording, one domain was targeting around 1,600 paid keywords, which reframes their organic numbers entirely.

Test the error rate once, on yourself

Here’s the only calibration exercise worth doing, and it takes five minutes.

Run Traffic Analytics on your own domain. Compare it to your analytics and Search Console. Write down the ratio. On some sites the estimate lands close, on others it’s out by half in either direction, and the direction is not consistent across markets.

That ratio is your error bar for every competitor number you’ll ever read in this tool. Without it you are quoting a figure whose accuracy you have never checked, about a site you have less data on than your own.

Two things gated behind money

Compare Domains is plan-dependent. In the recording it simply wasn’t available on the plan being used, so the walkthrough falls back to Semrush’s sample data to show what it would look like. Comparing four competitors side by side is the single most useful view in the tool and it is not universally included.

The full dataset is a separate subscription. Not a higher tier. A second product on top of your plan. With only the basic data, several reports inside the interface are visibly present and locked, Trending Pages among them.

That is two different money questions on one screen, and neither is obvious from a feature list. Both are worked through in what Semrush actually costs.

Use it for the comparison, never for the count

The honest workflow is narrow, and narrow is fine.

Good uses. Which of these five competitors is biggest. Whether a rival’s traffic is bought or earned. Which channel a competitor relies on that I’ve been ignoring. Which countries send them traffic. All of these are relative questions, and relative is what a model can do.

Bad uses. Reporting a competitor’s traffic as a number in a document. Sizing a market from it. Measuring your own site, where Search Console is free and actually counts things. Justifying a strategy on a figure you have never error-checked.

The workflow that actually pays is boringly simple. Pull five competitors, sort them by estimated visits, and ignore the numbers entirely. What you want is the ordering and the channel mix. If the biggest one earns 70 percent of its traffic organically and the third-biggest is buying half of theirs, those are two completely different companies to compete with, and no keyword report will tell you that apart.

Then check the country breakdown before you commit. I have watched people benchmark themselves for months against a rival whose traffic came almost entirely from a market they do not sell in. The estimate was wrong by some unknown margin and it did not matter, because the geography was the finding.

If you need to state a competitor’s traffic to someone, state it as a range with the tool named and the date attached, the same way you’d cite any estimate. Anyone who has run the calibration above will respect you more for it, and anyone who hasn’t will eventually get caught quoting a modelled number as a measured one.