Roughly three in ten people who visit your site never fire a marketing tag. GWI’s Q2 2025 data puts ad blocker usage at 29.5% of internet users worldwide, about 1.77 billion people, with the United States at 32.5%. Most blockers kill the request to connect.facebook.net or googletagmanager.com at the network level, so the loss is invisible. Nothing errors out. The events simply never exist.
I manage over $80K a month in ad spend and I deploy server-side Google Tag Manager and Meta’s Conversions API on client accounts. The question I always get is cost. My answer is that the cost conversation is backwards: you are already paying for the missing signal, every day, in the form of dumber bidding.
Three leaks in every client-side setup
Ad blockers are the blunt one. At the usage rates above, a meaningful slice of your conversions is invisible to the platforms that are supposed to bid toward them. The people who block ads still buy things. They just do it off the books.
Safari’s Intelligent Tracking Prevention is the subtle one. Since ITP 2.1, WebKit caps every cookie written through document.cookie at seven days. Your _ga and _fbp identifiers are set by JavaScript, so on Safari they evaporate in a week. A user who researches in week one and buys in week three becomes two different people in your reports. It gets worse. ITP 2.2 cuts that cap to 24 hours when the visit arrives through a decorated link, and a gclid or fbclid is exactly that. The visitors you paid the most to acquire get the shortest memory of all.
Consent friction is the legal one. In the EU, a declined banner means no tag, no cookie, no event. Even a well-designed banner loses a chunk of visitors before measurement starts.
Stack the three and the ad platforms are often optimizing on a fraction of reality. Smart Bidding and Advantage+ are prediction machines. Feed them less and you get worse predictions at the same price.
The organizational failure mode makes it worse. A broken tag gets noticed and fixed within a day. Thinner signal gets explained away as rising CPMs or seasonality for months, because nothing on any dashboard points at the tracking. I have watched teams do exactly that.
What the server-side stack recovers
Server-side GTM moves the tagging runtime out of the browser and onto a container you host on a subdomain of your own site. Two things change. Cookies can now be set in the HTTP response from a first-party endpoint, which takes them out of the document.cookie bucket that ITP caps at seven days, so identifiers survive. And the browser talks to your own domain instead of a blocklisted third-party host, which brings back a good share of the requests blockers were eating.
Meta’s Conversions API closes the loop on the Meta side: purchase and lead events travel from your server to Meta directly, deduplicated against whatever the pixel managed to send. Meta’s own product page claims advertisers saw an average 17.8% lower cost per result after setting up the Conversions API for web events.
Meta reports an average 17.8% reduction in cost per result for advertisers who add the Conversions API to their web pixel. That is Meta grading its own homework, but even a third of the claimed effect pays the hosting bill many times over.
The implementation detail that separates a clean setup from a double-counting mess is deduplication: pixel and server send the same event with the same event ID, and Meta keeps one. Skip that and your reported results improve for the worst possible reason.
One thing server-side tagging does not do is bypass consent. A declined banner still means no tracking, whether the tag fires in a browser or on a server. The stack repairs the technical losses, not the legal ones, and anyone selling it to you as a consent workaround should worry you.
My own experience is directional but consistent. An apparel e-commerce client moved to server-side GTM plus CAPI, and within days the event match quality on their purchase event climbed from the middle of Meta’s scale to the top band. Delivery steadied. The daily CPA chart stopped looking like an ECG. Nothing else about the account changed that month, which is as close to a controlled read as you get outside a proper lift test.
The actual cloud bill
Google’s own deployment guide prices a Cloud Run tagging server at approximately $45 per month per server and recommends running at least two instances so a single outage does not cost you data. Call the floor $90 a month, plus a little for logging. The same guide says an autoscaling fleet of 2 to 10 servers handles 35 to 350 requests per second, which covers most mid-size e-commerce sites with room to spare.
Managed hosting providers undercut Google’s list price, but even without them the math is not close. On a $10,000 monthly budget, $90 of infrastructure is 0.9% of spend. If restored signal improves your blended cost per acquisition by even one percent, the server is free. I have yet to migrate an account where the recovered events did not dwarf that cost.
Where I land
Server-side tagging is unglamorous plumbing, and for years I treated it as optional. I was wrong. Between blockers, ITP’s cookie caps and consent, a client-side-only setup now sees a partial view of your own customers, and every bidding algorithm downstream inherits the blindness. If you spend five figures a month, the hosting cost rounds to zero and the payback question answers itself. Put the container on a subdomain, deduplicate properly, and give the algorithms the data you already earned.