Privacy-First Telemetry Is a Downgrade You Should Choose Anyway
One analytics script, about 300 bytes after minification, and the meeting it started ran forty minutes past its slot. A consent banner was on the agenda, and halfway through someone landed the only question that mattered: name one decision the last quarter's dashboard changed. Nobody in the room could. That silence, not the banner, is the budget every telemetry decision in this post is argued against.
The claim: cookieless analytics without persistent identifiers is measurably worse at measurement, and it should still be your default, because the questions it deletes are questions a small product was never answering well. The rest of the post earns that sentence in three moves. First, the four events that survive a no-PII rule on one contact page. Then the questions that die when sessions die, named one by one. Then the two deployment shapes, self-hosted and cookieless hosted, and where their trade honestly sits.
Four events survive the no-PII rule on a single page
The design rule comes before any tooling: record actions, never people. On a contact page that rule leaves exactly four events standing: a page view, the form opening, the submit, and the submit failing. Four is not a compromise arrived at reluctantly. It is the number that answers the questions a consultancy site actually has: does the page get seen, do visitors reach the form, does the form work.
{
"name": "contact_form_submitted",
"ts_ms": 1749500000000,
"props": {
"path": "/contact",
"outcome": "ok"
}
}
The properties are the whole discipline. path is stored without its query string, so a campaign tag never lands in the event log beside a page name. outcome is one of ok, rejected, or api_error, a value your own validation code assigns, which keeps the validator's vocabulary rather than a stack trace out of the event log. There is no user id, no email, no free-text field, and no user agent beyond the browser family the analytics layer summarizes on the way in. Nothing in the schema could be joined back to a person even by someone staring at the raw table.
The visitor count deserves its mechanism written out, because "unique visitors" without cookies is an estimate, and the estimate has a declared shape. Plausible's documented approach hashes the visitor's IP, user agent, and the site's domain with a salt the server rotates daily, so the same person counts once within a day, and their two visits in different weeks stay unrelated by construction. Read that as the definition rather than a flaw: daily uniqueness, printed at the top of every dashboard.
Event names earn a paragraph of their own, because a schema drifts faster than its rules. We hold events to a naming convention of lowercase, underscore-separated strings such as contact_form_submitted, so a rename in code shows up in the log as a new name rather than a silent mutation of an old one, and renames happen on purpose. The list of every accepted event property lives in the repository, next to the validator that rejects anything unlisted. Both artifacts get reviewed like code, because they are the contract the analytics layer enforces on every future event somebody adds in a hurry.
The questions that die when sessions die
Sessions carry half of any dashboard's tabs, so the subtraction list is long and worth reading slowly. Gone: funnels that span more than a day, return-visitor cohorts, and anything that joins one person to their previous visit. The tools that stitched sessions together got their thread from a cookie or from local storage, and both are precisely what the no-PII rule refuses.
What remains covers the census questions. Per-day uniques, per-day counts of contact_form_submitted, the share of opens that convert, and top referrers for first touch all survive intact. Those four lines answer the question a meeting can act on this month: is the site working, yes or no, on the exact pages it ships.
The loss concentrates into one sentence: cookieless telemetry cannot tell you whether people come back. Retention, the metric every growth pitch treats as the heartbeat, is unavailable by design, and no config flag returns it. You can buy a piece of it back with outcomes instead of returns, because a returning visitor who never fills the form is invisible to the analytics layer and fully visible to the inbox. The second email is the cohort analysis.
Self-hosted or cookieless hosted: pick your owner
Both deployment shapes clear the no-cookies bar. Where they differ is in who holds the data and who holds the pager. Self-hosted options such as Umami, Plausible Community Edition, or Matomo configured without its cookies, which is not Matomo's default, put the events table inside your own infrastructure and bill you in maintenance: one database that only grows, plus backups, plus upgrades on a Friday you did not plan. Hosted cookieless options such as Plausible, Fathom, or Cloudflare Web Analytics remove the pager entirely and answer the consent question the way their products are built to: no cookies, no banner, with your own counsel holding the final word on the ePrivacy reading.
Two constraints pick between them, and the decision transfers to any team of your size. If the telemetry exists to tell four people whether a marketing site works, hosted wins, because running a self-hosted analytics box costs more every month than the marginal insight it returns on a site this small. If your events are a product surface, feeding dashboards or experiments that customers touch, self-hosted wins, because the alternative means routing your users' activity through someone else's billing relationship. Cloudflare's free tier sits apart and is worth naming plainly: it counts pages, referrers, and Core Web Vitals, then stops. That is either the whole job or obviously short of it, and both answers are legitimate.
The session we tried to rebuild, then deleted
Here is the dead end we would rather confess than repeat. Our first schema carried a visitor_token property, a random id written to local storage, so we could stitch a contact visit to its second visit weeks later. It stitched beautifully for one afternoon. Two code reviews later we deleted it, after agreeing in plain words that local storage holding a persistent identifier is a cookie wearing different syntax, and it drags the consent machinery back in through the window you left open. The salt-rotating daily hash does the privacy work precisely because it forgets, and forgetting is the feature we were trying to undo.
What this telemetry cannot tell you
The limits stay listed, because a downgrade you choose only stays a good choice while you remember what you gave up. No cross-day funnels. No return-visit cohorts. No durable device profiles, so a person on a laptop and on a phone counts as two uniques forever. Unique-visitor numbers drift by a few percent depending on how shared IPs hash, which is why we treat a 3 percent week-over-week move in uniques as noise at this site's traffic. Anyone who needs session-level truth, say for debugging a leaky funnel in a product, needs consent-based analytics with a banner and a data-processing agreement, and that is a different post with a different trade.
One decision closes the argument. If your dashboard's most useful page is the raw event list, four events and a cookieless hosted counter will tell you everything the dashboard was going to change, and you can delete the banner. If someone asks for a funnel, ask them which decision it changed last quarter, and hold that standard until they name one.
AngoraTek builds sites and products that pass a privacy review without a consent banner. If your dashboard owes a teammate a decision it has never paid, bring us the four events it should be answering instead.Thanks for reading this far. If this sounds like the problem you're wrestling with, tell us about it. The team reads every message.
Start a conversation