GA4 you can actually make decisions with.

Four things: a measurement plan that names what matters, collection you can trust, history that outlives the interface's retention window, and a handful of reports that end in decisions. In practice the first two carry the value, because clever analysis on duplicated events produces confident nonsense. A good engagement leaves you with a documented property, conversions that reconcile to your sales system, raw data in your own cloud project, and a team that can extend all of it without calling us.

Neither on its own, and the gap is normal rather than a fault. Ad platforms count conversions they believe they influenced, each using its own attribution window and modelling, so the totals across platforms will always exceed reality — every platform claims the same sale. GA4 counts by its own rules and session logic. The sales system knows what closed but little about the touches before it. We use each for what it is good at, reconcile them on a schedule, and document the expected gap so it becomes a known quantity instead of a monthly surprise.

On a standard property, Google keeps user and event data for a maximum of 14 months, and the shorter option is the default many properties are still running. That is the single most common irreversible loss we find. The answer is the BigQuery export, where the daily job is capped at a million events per day and the streaming export has no cap. Switch it on and your history accumulates in your own project, unsampled and permanent. It is inexpensive at mid-market volumes and the one decision that cannot be backdated, so we treat it as week-one work.

The interface is fine for behavioural analysis and day-to-day reporting when it is configured well, and plenty of clients stay there for a while. You need the warehouse when revenue truth lives in a sales or billing system, when refunds and cancellations change the picture materially, when sales cycles run for months, when reports start hitting sampling or cardinality limits, or when you need history beyond the retention window. Because the export only collects from the day you enable it, we usually enable it immediately and build on it later — the cost of waiting is data you can never get back.

It is, for critical conversions and for any site where measurement materially affects bidding. Google's documentation explains the mechanism: in a server-side setup the browser sends one request per event and your own container fans out the vendor calls, which moves work off the page and keeps collection under your control. The practical gains are more durable conversion data, less exposure to browser tracking prevention, and fewer third-party scripts competing with your page. It adds a cloud cost and a maintenance responsibility, so we scope it where the decisions justify it rather than by default.

Substantially, which is why we implement it carefully rather than treating it as a banner setting. Google's own guidance notes that consented users are typically 2-5x more likely to convert than unconsented users, so a broken signal does not just remove a slice of data, it distorts the shape of what remains. Behavioural modelling fills part of the gap by estimating declined users' behaviour, and it is genuinely useful for trends — but it is an estimate, so we never reconcile modelled figures to the cent and we always report what was observed alongside what was modelled.

Usually, yes, and that is the outcome we aim for. Most rushed migrations are structurally sound and simply carry a handful of expensive faults: duplicated events, conversions defined on page views, a consent banner that blocks the tag entirely, ad platform links pointing at the wrong property. We repair those in place, keep the historical data intact, and only recommend a fresh property when the event naming has become genuinely unrecoverable — which is rare. When a new property is the right call, we run both in parallel so you never lose continuity while the replacement builds up history.

Less, in our experience. Most of the effort teams currently spend goes on reconciling exports and re-explaining why two reports disagree, and that work largely disappears once the definitions are agreed and the joins happen automatically. We also build fewer dashboards than clients expect, because three views people trust beat fifteen nobody opens, and every extra chart is a maintenance liability. What we do ask for is about an hour a week from someone who owns the numbers internally — measurement with no internal owner decays within a couple of quarters.

The first engagement is a fixed-scope audit and rebuild — audit, remediation, measurement plan, BigQuery export and a baseline set of reports — priced up front so you approve a number rather than an open-ended retainer. Ongoing support is a monthly fee scoped to the reviews, testing and new tracking you want, and it can be paused when there is nothing to do. We do not price analytics as a percentage of media spend: the two are unrelated, and tying them creates exactly the wrong incentive.

We can, and many clients prefer it because the loop closes faster when the team reading the data is the team changing the campaigns. Our paid search, paid social and SEO teams work from the same property and definitions. It is equally fine to buy the measurement work alone while other agencies buy your media — we report on their performance honestly and share the methodology with them, because analytics only one agency can interpret is not much of an analytics setup.