The Super Data Collectors

How does someone know how many visitors a third-party website has?

Every second pitch deck contains a Similarweb number. Competitive analyses, due diligence, media planning, hedge fund models: somewhere in there sits a bar chart with monthly visits to a website to which the author has no access whatsoever. And it works surprisingly well.

The question behind it is a nice technical question, and it has a clean answer. The number is not a measurement, it is an extrapolation. Just like an election forecast: a sample is observed, corrected, calibrated against known truths, and then stretched to the total population.

Anyone who understands an election forecast understands Similarweb. It's the same four work steps, the same sources of error, the same kind of caution when reading the result. The only difference is the sample: instead of a thousand people called, it's millions of browsers.

This report goes through the machine from the beginning. First the four inputs, as the company itself describes them. Then the calculation method. Then the origin of the panel, which is its own and quite interesting story. After that the company, the numbers, and the accuracy.

First the scale: Similarweb is not a special case in this industry, but the best-documented example. What is written here applies in principle to every provider of market data from behavioral measurement.

Millions of people contribute measurements without that being their goal

Now for the most difficult ingredient. A panel of this size doesn't arise from people voluntarily signing up for market research. It arises from a product being useful and recording on the side.

This isn't a Similarweb invention, it's the basic model of behavioral measurement since TV ratings have existed. What's new is the scale: instead of a few thousand households, it's millions of devices, and the compensation isn't cash but a free tool.

The origin of this panel has been documented over ten years by security researchers, universities, and an ad blocker manufacturer. The following timeline is the sober summary of these works.

The answer is in the terms, and it is surprisingly open

For the question of what data actually flows, you don't need a security researcher. You just have to read the right file, and that's the small difficulty: Similarweb's privacy policy explicitly does not apply to the browser extension. It says so in its own sentence.

The declaration that applies to the extension is on a different domain: similarsites.com. There it is stated very clearly who is responsible and what is collected. This is not a leak and not a discovery, this is the text that is agreed to upon installation.

The last line in the table is the most economically important: the result may be passed on to business customers. This is exactly the product that Similarweb sells.

From jewelry store to stock exchange: Or Offer

Or Offer, born in 1983, grows up in Tzur Hadassah near Jerusalem, as a computer kid with parents who design jewelry and struggle with the uncertainty of this profession. After school, intelligence unit 8200 wants him as a programmer. He declines and goes to Oketz, the canine unit, among other things to a post in the Gaza Strip. About this time he later says it taught him to stand up to experienced people as a young person.

Afterwards he studies business administration part-time at IDC Herzliya and builds two jewelry stores, in Zichron Ya'akov and Netanya, plus an online shop. The chain is called Maya Offer, after his mother. His parents push him towards high-tech.

The trigger is a name: a supplier mentions the designer David Yurman, who combines stones with silver. Offer wants to find comparable products online and realizes that there is no tool for this. Similarweb is created in 2007 as a browser extension that suggests similar websites to a website. Hence the name, and hence also the panel: data collection is originally the byproduct of a search function.

He is 23 and works for over a year without salary. In 2007 an important collaborator leaves the project because he doesn't believe in the product. At 24 Offer thinks about quitting and stays: "I had nothing else to do, so I continued."

In 2009 Yossi Vardi gives him one million dollars, for about five percent. The company lives on this for five years, with seven or eight people. In 2011 comes the pivot that decides everything: away from the consumer product, towards an analysis platform for companies and investors. The consumer product doesn't disappear in the process, it changes roles. From now on it is no longer the product, it is the measuring point.

Offer was sole founder and considers this an advantage: "When there are multiple founders, each needs a role to feel important. Precisely because I was alone, that saved me a lot of politics."

In 2014 Naspers leads a round of 18 million dollars. In May 2021 the company goes public on the New York Stock Exchange, valued at 1.6 billion dollars, and raises around 180 million dollars. Headquarters is still Givatayim near Tel Aviv, plus New York and London.

Four sentences to get the picture right

First: nothing is being accused here. There is no case in which a legal violation has been proven to an authority. This report describes a mechanism, it does not accuse.

Second: Similarweb contradicts the 2016 presentation. The company has stated that these are "repeated foreign reports from 2016 that were already baseless at the time and were refuted many years ago." This sentence stands here as it was made.

Third: the company has visibly restructured. After 2018, away from dependence on the Store, toward purchased partner data, with explicit exclusion of Stylish data in the Check Point agreement and a consent dialog for AI content since January 2026.

Fourth: this is an industry model, not an isolated case. Behavioral panels from consumer products have existed since television ratings. The 2026 scan found 287 extensions, not four. Similarweb is the best-documented example here because it is publicly traded and therefore must be accountable.

And the point at which the whole thing dissolves: there is no method to estimate third-party traffic without observing someone somewhere. The alternative to a panel is not a better panel, but no number at all. Anyone who uses the number uses the method with it.