Data integrity first, always
Rule out the tracking fault before you spend an hour on the market. It is eliminated first because a data fault makes every downstream number meaningless, and most weeks that is the whole answer in seconds.
For in-house teams
Get the investigation done before Monday
You can already read a chart. What you cannot get back is the half day spent proving the drop was a broken tag rather than the market.
Key takeaway
You are not short of dashboards. You are short of the 30 to 60 minutes an anomaly costs before you can even say what it was, several times a week, always on the day you had something else planned. That checklist is worth automating. The judgement at the end of it is not, which is why every verdict comes back to you.
The triage
The boring half of the job, done for you
When ROAS moves, the first hour is not analysis. It is elimination: was it the pixel, the spend, the creative, the funnel, the UTM labelling, or the market. Six checks that have to be run in order and that nobody enjoys running.
A confidence layer, stated
Know how far to trust the answer before you act. Deterministic for a data fault, data-backed when internal evidence was found, contextual when internal causes were ruled out, labelled rather than blended into one confident paragraph.
It admits when it is guessing
See when the run had nothing to go on. Each investigation records whether a known pattern matched or it could only report facts, because a fluent narrative over no evidence is how these tools usually fail.
The audit
Numbers you can check rather than trust
You are the person who gets asked where a number came from, and "the tool said so" is not an answer you can give twice. Everything here is built to be checked.
Thresholds printed, not hidden
Defend the number when someone asks where it came from. Page audit bands ship with a floor and a ceiling as numbers, labelled a documented first pass, because a threshold you can argue with beats a grade you cannot.
Six models on the same window
Show a stakeholder what platform bias costs, in one screen. Run last-click against data-driven attribution on the same conversions and read the gap yourself.
What we refuse to claim
Find our limits stated rather than buried. No citation-to-revenue bridge and no lift figure from an experiment nobody has run, listed on the pages they apply to, because a vendor's limits are what you most need to know.
The exit
The raw events are yours to take
Being technically capable means eventually wanting the data somewhere else, whether that is a notebook, a warehouse or the next tool. A platform that makes that hard has made a decision about you.
Your bucket, your format
Take the raw rows wherever you want them. Events and attribution tables to Google Cloud Storage or Amazon S3, in Parquet or CSV, on the same schema the interface reads.
Full history, no retention cliff
Query a year-old cohort without asking anyone. Nothing is aggregated away at a plan boundary, which is the quiet way most tools make leaving expensive.
Query it from your own tools
Ask a question without rebuilding a view to answer it. The MCP server (Alpha) connects Claude, ChatGPT or Cursor to the same deduplicated numbers your team reads.
FAQ
Direct answers.
GA4 tells you what happened on the site and takes no position on whether your ad platforms are over-claiming. A BI tool visualises whatever you load into it and takes no position on whether the numbers are right. The layer missing between them is a single conversion definition, reconciled against your real orders, that every channel is measured on. That, plus the tracking audits that tell you when the collection underneath has quietly broken.
That is the design goal. Thresholds are published as numbers, attribution models can be run side by side so the gap between last-click and data-driven is visible rather than asserted, and every investigation states its confidence layer and whether it matched a known pattern or only reported facts. Where a claim cannot be made honestly, the page says so instead of hedging.
One snippet in the head for collection, five minutes. Connectors are OAuth, a couple of clicks each, with historical data loading on first connection. Custom events and server-side forwarding take about half an hour with the documentation open. There is no data team requirement and no warehouse prerequisite, which is the point: the whole thing is aimed at a capable person who does not have three months.
Yes, and while you are a customer rather than only on the way out. Raw events and attribution tables export to your own Google Cloud Storage or Amazon S3 bucket in Parquet or CSV, on the same schema the interface reads, with no retention limit deleting history first. A data-ownership claim you can only test at cancellation is not a claim.
Possibly, and it depends on your spend. Below roughly 50,000 EUR a month in ad spend, a cheaper single-purpose tool is usually the right buy and we would rather tell you that than sell you a platform you will underuse. Above it, the reconciliation problem and the time cost of anomaly triage are usually worth more than the subscription.
14-day free trial
Get the half day back
14-day free trial. Tracking audits, anomaly investigations and raw export are on every plan, including Measure.