How to Use Similarweb for Competitor Analysis

Similarweb dates its technology to 2011, and its annual report and its own documentation describe the same four classes of input. Website and app owners share their first-party analytics directly. A contributory network of consumer products collects de-identified device behaviour at the site level. An automated index captures publicly available pages every month, alongside census data such as country population. And a set of data partnerships supplies pre-analysed signals from internet operators, measurement firms and demand-side platforms. Those four streams are cleaned, blended and run through predictive models, and what reaches you is the output of that model rather than a count.

Three of those four inputs describe watching from outside. The first does not, and it is the reason this product behaves differently from every other traffic estimator. Companies register, indicate whether they are willing to share their own analytics, and some of them choose to publish what they share. When that happens the headline figure on the profile stops being a model of the company and becomes a number the company handed over. Working out which of the two you are looking at is the first move in any competitive read here, because it decides whether you may quote a level or only an ordering.

What it never does is count in real time. The documentation is direct about it: monthly figures publish by the tenth of the following month, daily figures land about seventy-two hours after the day ends, and the platform states plainly that live traffic is not available because it measures actual user behaviour rather than live traffic. Two smaller exclusions are worth knowing before a number surprises you. Connected television is outside the dataset entirely. And the whole platform, including its interface and its feeds, runs on an Eastern Standard clock that ignores daylight saving, so a European or Asian rival’s daily figures are attributed several hours away from that market’s own calendar.

1. 1Open the free profile before you open anything you pay for. Type the competitor's domain into the public traffic checker, or click the extension while you are on their site. Both are free, neither needs an account, and between them they answer the visit level, the engagement metrics, the channel mix, the country split and the ranking. Most competitive questions stop here, and knowing that saves you from buying a subscription to answer them.
2. 2Establish whether you are reading an estimate or the owner's own analytics. A company can connect its analytics and choose to publish the connection, in which case the visit figure on its profile is the session count its own analytics recorded. That is a different kind of evidence from a model, and it changes what you may say about the number. Check for the verification marker before you quote anything, and record which kind you found.
3. 3Place the domain inside the two published thresholds. Below five thousand visits in the most recent month nothing displays at all. Above a hundred thousand the vendor states it is more confident in its estimates. In between, a figure appears without that statement attached. Write the band next to the number, because the band decides whether you may report a level or only an ordering.
4. 4Fix the country, device, period and subdomain handling, then keep them fixed. Every figure is a function of four filters, and subdomains are included unless you exclude them, so a rival whose product sits on an app subdomain is being counted differently from one whose product sits behind a separate domain. Choose the settings once for the whole competitive set and write them into the sheet beside the numbers.
5. 5Build the set from Similar Sites, then reconcile it against deals you actually lost. The similarity list is assembled from audience overlap, which surfaces content sites and marketplaces that take clicks from you and misses rivals you meet only inside a procurement process. Take the list as candidates, cross it against the names your sales team reports, and keep the union rather than either one alone.
6. 6Record each figure with its filters and the date you read it. Store the number, the four filters, whether the profile was verified and the day you looked. When you refresh the series, pull every month again rather than appending one row, because the vendor restates history when it changes its methodology and an appended sheet quietly mixes two versions of the same past.

A site owner can connect their analytics to Similarweb and choose whether that connection is public or private. The support documentation states the consequence in one sentence: once a website is verified, the platform shows the number of visits as they appear in Google Analytics, and when the connection is public that figure is visible to anyone searching for the site. So a verified profile is not a better estimate. It is not an estimate.

The important half of that is what verification does not touch. Similarweb states separately that its rankings are built from its own estimations and that connecting analytics will not change a site’s rank. Neither does it convert the channel mix, the country split or the engagement metrics, all of which stay modelled. A verified profile therefore carries a measured visit count sitting next to a modelled rank and a modelled composition, on the same screen, in the same typeface, with nothing to tell them apart.

Knowing who verifies, and why, tells you how often to expect it. Similarweb’s pitch to site owners is explicit about the audience: verify and get discovered by advertisers so you can monetise your traffic, put your best foot forward for investors, and help partners see your reach. Its own methodology page says that companies which monetise traffic, giving publishers as the example, often choose to share publicly. Verification is a marketing act, and it pays off for businesses whose product is attention.

The consequence for business software research is unhappy and worth stating plainly. A private company selling seats has nothing to gain from publishing its traffic and something to lose, so the rival you would most like a real number for is the one least likely to have supplied one. Expect an estimate. When you do find a verified competitor, treat it as a calibration point rather than a curiosity: you now have one domain in your category where the true level and the modelled rank sit side by side, which is worth more than any global accuracy average.

The same mechanism runs in the other direction, and this is the one number on the platform you control. Connecting publicly replaces the estimate of your traffic with your real session count for every person who looks you up, competitors included. That is a deliberate trade rather than a hygiene task, and the private option exists precisely for the case where you want your own team calibrated without publishing the answer.

There is peer-reviewed evidence on this rather than only vendor material, and it points somewhere counter-intuitive. Jansen, Jung and Salminen published a comparison in PLOS ONE on 27 May 2022 that set Similarweb against the site owners’ own Google Analytics for 86 websites drawn from 26 countries and 19 verticals, using twelve months of monthly averaged data. They reported both how closely the two rankings agreed and how far apart the levels sat.

Read the two columns together. On the volume metrics the ordering is close to perfect: put a set of competitors in size order using Similarweb and the ordering will usually be right, even though the levels sit well below what the sites themselves recorded. On the engagement metrics the agreement drops to moderate, roughly half the strength of the volume figures, and the error runs in a consistent direction, with bounce rates reading high and sessions reading long.

Two caveats belong with those numbers. The data runs from September 2019 to August 2020, so the products have moved since and the percentages are period-specific even though the pattern is not. And the authors drew their sample from a public list of large domains, keeping sites where Similarweb had analytics available or where the owner had already connected their own, which means part of the comparison may be against profiles that were verified rather than purely modelled. That is a reason to lean on the correlations rather than on the gaps, and a second reminder that the two kinds of profile need separating before anything is concluded.

Site size then decides how much of any of this you may say out loud, and Similarweb publishes its own thresholds for that. Below five thousand visits in the most recent monthly snapshot it displays nothing at all, showing “not enough data” and stating that the site is monitored but the estimate would carry too much variance. Above a hundred thousand monthly visits it states that it is more confident in its estimations. Between those two lines a figure appears with no such statement attached to it, which is exactly where most private business software companies live.

Two details make the floor sharper than it first looks. It reads the latest snapshot rather than an average, so a competitor with a seasonal collapse can disappear from the platform on the strength of one bad month. And features that only carry desktop data apply the floor to desktop alone, so a rival with a mobile-heavy audience can show healthy totals in one view and nothing at all in another. Similarweb also cites an independent comparison on its own accuracy page and names five thousand to a hundred thousand monthly users as the band where it performs best. The two statements are about different things, one being its own confidence and the other being how it placed against rival estimators, and in practice they point the same way: state the band beside the number. It is also the reason to settle on one estimator for the year, because two vendors modelling competitor website traffic are producing two different scales rather than two readings of one.

Most people assume a traffic history accumulates: last year’s months stay where they were and new ones arrive on the end. That is not how this dataset works, and the vendor documents the alternative in two independent places.

The first is a release note. A new data version launched on 28 July 2024, and its headline was not the new coverage but the rerun: five full years of history recalculated for total visits, unique visitors, page views, pages per visit, bounce rate, average visit duration, rank, device share, geography, deduplicated audience and marketing channels. Eleven metrics, which is essentially everything a competitive researcher uses. The same note adds more than thirty million domains, moves the learning set onto Google Analytics 4, improves filtering of automated traffic, rebuilds the country models for several large markets, and states that the update reaches the platform, the interface, the feeds and the free tools together. It also says in plain terms that some sites would see their numbers rise or fall.

The second is the company’s annual report, which carries a risk factor headed by the warning that the business may be harmed if it changes its methodologies or the scope of what it collects. The text tells investors that it has changed its collection and aggregation methods before and may again, and that estimates for future periods may end up incompatible with the estimates it published for prior periods. That is a listed company telling its shareholders that the back series is not guaranteed to match itself.

Against that, the support documentation answers “why do I see different data for a month I already checked” by pointing at filters, and it is usually right: a different country, a different device setting or a different period will move a number without any model changing. Both things are true, which is what makes this awkward in practice. The common cause is your own settings and the uncommon cause is a restatement, and a bare number in a spreadsheet cannot distinguish between them.

The getting-started documentation defines the term for you, and the definition is honest enough that it should change how you use the output. In the digital world, it says, a competitor is any site that pulls clicks away from your site, and it warns that your digital traffic competitors may look a little different from your well-known business competitors. That is a definition built around attention, not around revenue.

The Similar Sites feature applies it. It returns up to forty domains with a similarity score attached to each, assembled from audience overlap rather than from anybody’s opinion, which is genuinely useful and genuinely not the list you present to a board. It reliably surfaces the review sites, comparison sites, marketplaces and publications standing between you and your buyers, and it will occasionally hand you a rival nobody internally had named. It will just as reliably miss a competitor whose marketing site is thin because they sell through a field team, one whose product lives entirely behind a login on another domain, and one you only ever encounter halfway through a procurement process.

The reconciliation is not difficult and almost nobody does it. Take the overlap list as candidates, put it beside the names your own sellers report from lost deals, and keep the union. Names on both lists are your real competitive set. Names only in the overlap data are competing for attention and belong in content and search planning. Names only in the deal data are competing for money and will never appear in a traffic tool at all, which is the strongest argument for treating win/loss interviews as a data source rather than as a post-mortem ritual.

One practical constraint shapes the whole exercise. The platform compares five domains at once, meaning yours plus four, and the free public checker allows one comparison. That ceiling is low enough that the set has to be chosen rather than assembled, and stable enough that changing who is in it destroys the comparability of everything you recorded before. Pick the four rivals who share buyers with you rather than the four largest names in the category, and run a second set separately if your market genuinely splits in two.

The free layer is wider than the pricing page implies, it has no expiry, and an account that finishes a trial reverts to it rather than closing. Each of these answers a different question, and the one most people skip is the one that gives the most away.

The public website profile

Enter a domain in the traffic checker and the page returns a monthly visit estimate, the trend behind it, engagement metrics, the channel breakdown, top countries and a similar-sites list, without asking anyone to register. Two documented limits shape what it is good for: it compares your domain against one other rather than four, and it reports all traffic without the desktop and mobile split, which is a paid view. Use it for a quick read on a single rival and for checking whether a domain clears the display floor at all.

The browser extension

This is the surface worth knowing about. It installs on Chrome, Edge, Firefox and Opera, needs no account whatsoever, and returns for any page you are looking at: visits over time, bounce rate, pages per session and average visit duration, the full channel breakdown across direct, organic, paid search, social, email, referrals and display, the country split as percentages, the global, country and category rankings, the keywords with estimated click volume and cost per click, and the share of traffic arriving from AI assistants. That is more than the free website page gives, from a surface nobody has to register for.

Top Websites and the category rankings

A monthly ordering of domains by the rank algorithm, filterable across more than 250 categories and viewable for all traffic, desktop or mobile. The table carries traffic share, the month-on-month change, the rank, monthly visits, the device split, visit duration, pages per visit, bounce rate and whether the domain runs Google’s ad network. Alongside it sit trending lists, a mobile-against-desktop view and an annual hundred-domain round-up. For category sizing this beats looking up rivals one at a time, with the caveat that the ranking counts a domain together with all of its subdomains.

The Company Profile Checker

Domain lookups break on companies that run several brands, and this is the fix. Search a company rather than a URL and it returns the portfolio of sites it operates, its subsidiaries, its audience interests, more than three years of history and, where the company is listed, its ticker performance beside the digital figures. For any rival that acquired its way into your category, the single-domain view has been understating them and this is where that becomes visible.

The AI Traffic Checker

Free, and the newest thing here worth a researcher’s attention. It reports how much of a domain’s traffic arrives from assistants such as ChatGPT, Gemini, Claude and Perplexity, which of its pages receive that traffic, and the prompts associated with it, across up to five competitors at once. Treat the share as small and the composition as the finding: which of a rival’s pages assistants have decided to cite is a ranking you can act on this quarter.

The Website Technology Checker

This one runs the lookup backwards. Instead of asking what a site is built with, you name a technology and get its usage across the web, its top industries and countries, and lists of the sites running it. That answers a question a per-site checker cannot: how much of your category has already adopted the thing your product integrates with. Going the other way, from one named rival to their stack, is a different job with different instruments: a competitor’s website technology is read from response headers, a local fingerprint and a dated adoption archive, none of which is this tool.

The keyword generator and the search volatility tracker

The keyword tool builds lists from a seed term with search statistics for Google, YouTube and Amazon, filterable by country and intent, and the volatility tracker reports daily movement in Google results so a rival’s sudden traffic change can be checked against whether the whole category moved. Neither replaces a dedicated search platform, and both are enough to sanity-check a story before you escalate it. What neither will tell you is which of a rival’s competitor keywords were won deliberately and which they rank for by accident, and that distinction is the whole point of the exercise.

Free app analytics and the app rankings

A web-only read badly understates any competitor whose customers live in an app. The free app analytics and the published Android, iOS and marketplace rankings give downloads, category position and engagement signals, which is the correction. Where the question is what users actually say about that app rather than how many there are, app store reviews carry the text.

The artefact this produces is not a traffic number. It is a row that can still be defended in a year, which means the metadata beside the figure matters as much as the figure.

The columns that make the sheet defensible

Eleven columns cover it: domain, month, visits, whether the profile was verified or estimated, the size band the domain fell into, the rank, then the four filters that produced the number, being country, device, period and whether subdomains were included, and finally the date you read it. The last five are the ones people leave out and the ones that decide whether next quarter’s comparison is real. Add a twelfth column for notes the first time a rival changes domain, because that event resets their record rather than carrying it over.

Three routes out, and what each one costs

By hand is perfectly respectable for five competitors and one number a month, and the extension makes it fast. Beyond that the platform’s exporter and its interfaces share one currency called data credits, and the rule is that credits are consumed by results returned, whether you download a report or call the interface, so a download is metered just as a programmatic pull is. The published example makes the arithmetic concrete: ten domains against five metrics across twelve months in two countries costs 1,200 credits. The cost of a competitive question is therefore something you can work out before you ask it, which is unusual and genuinely useful for budgeting a quarterly process. The standard interface allows ten requests a second across roughly one to fifty domains at a time, the large-volume interface reaches a million domains and five years of history and delivers into a warehouse, and access needs its own entitlement on top of a subscription.

Two quirks that will confuse the sheet

A spreadsheet add-in for Excel, an add-on for Google Sheets and a connector for Looker Studio all exist, which means a non-technical team can refresh a competitive tab without anyone writing code, and that is usually the right answer for a marketing team. Watch one detail when you do: the interface rounds figures to whole numbers in the platform but returns fractions to spreadsheets and feeds, so an export arrives looking more precise than the platform that produced it. The decimals are an artefact of the pipeline, not a finer measurement, and they should be rounded away before anyone sees them.

Three tiers of access exist and they are easy to confuse, because the vendor uses the word free for two different things.

Free, permanently, without an account

The browser extension, the public website profile, the technology checker, the company profile checker, the AI traffic checker, the keyword generator, the search volatility tracker and the published rankings all sit here. Similarweb describes this layer in its own words as a high-level sample showing a limited set of features without deeper drill-downs, which is fair. Somebody investigating a handful of named rivals, rather than assembling a prospect list, gets a long way here without ever paying.

The trial, and what happens when it ends

Registration with a work email opens the full platform for a trial period. Cancel before it ends and there is no charge; let it run and the selected package bills. Either way the account does not close: it reverts to the permanently free tools listed above, and a cancelled paid subscription does the same. That matters for planning, because it means losing budget does not mean losing the extension.

What a subscription actually decides

Four things, and none of them is access to the basic numbers. Depth of history is sold separately, with the traffic endpoints documenting 37 months subject to your package. Country coverage is chosen rather than included, from more than 190 available. Daily granularity, the desktop and mobile split, and the deduplicated cross-device audience figure are gated. And the exporter and interfaces come with a credit allowance that typically resets monthly and is not refundable, which admins allocate per tool and per person. The published plans split between a self-service family aimed at individuals and small teams and a contracted family for larger organisations; prices are listed but change often enough that the caps and the gating are the durable part to plan around.

The last two rows are the ones worth acting on. A traffic estimator can tell you that something moved and never why, so a rise on its own is a prompt to look at what a rival shipped, priced or promoted in the weeks before it. Paid movement in particular is better read from the creative itself, which is why competitor ad spend works from ad archives rather than from an estimator’s percentage.

Reading a competitor’s public profile raises nothing: the vendor sells that analysis as its product and the company being researched has no say in it. The questions worth asking on this page point the other way, at what your own use of the platform discloses.

• Connecting your analytics publicly is a disclosure decision, not an admin task. It replaces the estimate of your traffic with your real session count for every visitor to your profile, and the vendor markets that reach in the tens of millions. Whoever signs off on what your company says about its size should be the person who decides this, and the private option exists for teams that want the calibration without the publication.
• The consumer products are collection instruments as well as viewers. The company’s annual report describes a contributory network gathering de-identified behaviour through products distributed on the major app stores, and names its own browser extension and mobile app among them. Nothing about that is hidden and the documentation states that data is aggregated at site level, carries no personally identifying information and uses no address data. It is still a question your security team would rather answer before an extension is rolled out across a research team than after.
• There is a published route out, and it applies to you as an individual. The same report records that the company determined it must register as a data broker in some United States jurisdictions and that it honours opt-out requests. If someone on the team objects to being part of the measurement while they research it, that is an available answer rather than a reason to avoid the tool.

• Money. Traffic is not revenue, and two sites with identical visit counts can differ by an order of magnitude in what they earn from them. Proxy: entry price and packaging on the rival’s own pricing page, which is published and exact, against the traffic as a rough demand signal.
• Named visitors. The dataset is aggregated at site level by design and carries no personally identifying information, so no amount of drilling produces an account name. Proxy: the evidence companies publish about their own vendors, which is a different exercise with different sources.
• What happens after the login screen. Product usage, retention and activation live behind authentication and often on a different domain. Proxy: the pattern of a rival’s application subdomain against their marketing domain, plus what their engineering and support job adverts describe.
• A small or newly launched competitor. Under the display floor there is nothing, and a domain added during a coverage expansion begins its record at the date it was added rather than at its real beginning. Proxy: the absence itself, treated as a weak upper bound, plus branded search interest and indexed page count.
• Why the line moved. An estimator reports the effect and never the cause, which is the limitation that matters most in practice because it is the cause you would act on. Proxy: a dated record of what the competitor shipped, priced and published in the weeks before the move.

The publication calendar sets the cadence for you. Monthly figures arrive by the tenth of the following month and daily figures about seventy-two hours after the day closes, so a monthly refresh scheduled after the tenth is the natural rhythm and anything more frequent re-reads the same numbers. The rank in particular moves once when the monthly data lands and then holds for the rest of the month, so checking it weekly returns the same value four times.

Four events justify an off-cycle look. A rival changing domain, because their record starts again rather than carrying over. A brand or product launch, where you want the baseline captured before the campaign rather than after. An acquisition, where the company view rather than the domain view is now the honest one. And a published data version update, which is the moment to re-pull the whole series rather than trust the rows you already have.

A competitor tracking spreadsheet with the four filter columns beside each figure is sufficient structure for a set of five rivals, and it is the thing that makes the quarterly comparison honest rather than approximately remembered.

Work out the arithmetic of the calendar and the problem becomes hard to unsee. A rival changes their pricing page on 3 September. The visits that change causes are collected through September and published on 10 October. Somebody opens the file in the middle of October, notices the line moved, and spends a week finding out why. Six weeks have passed between the move and the explanation, and none of that delay came from anyone being slow. It is the publication schedule, and checking more often does not shorten it by a day. The daily view helps at the margin and arrives three days late with enough noise at the domain level that a single day rarely means anything on its own.

Six weeks is roughly the length of a sales cycle stage. It is long enough for a rival to brief their field team, run the campaign, and start appearing in your deals with a story nobody on your side has heard yet. The cost is not that you missed a number; it is that the first person in your company to learn about the change is a seller losing to it.

This is the gap competitive intelligence platforms exist to close, because they watch the cause rather than the consequence. Flares tracks what a competitor publishes, prices, ships and says about itself, and dates each change, so the pricing move surfaces the week it happens instead of being inferred from a line six weeks later. No product in this category can hand you the visit count sitting inside a rival’s own analytics. That figure lives on their side of the login and nowhere else, and an estimate stays an estimate however early it reaches you.