Method
How I work
Every consultant says they are data-driven. Here is what that means in practice for me, with examples from real engagements. Figures are anonymised.
1. I pull the data myself
I do not wait for decks or exports. I work directly in Google Ads, GA4, Tag Manager, BigQuery and the platform APIs, and when the built-in reports cannot answer the question, I write a script that can. My standard pattern for any investigation is a Google Ads script and an analytics script, each writing to its own sheet with an emailed summary, so the evidence is repeatable and anyone can rerun it. If a number matters, I want to be one step from the raw data, not four steps and two people removed from it.
2. Headline numbers get decomposed before they get believed
Platform figures are a starting point, not an answer. On one account, a reported six thousand conversions in a week decomposed to roughly 120 real sales; the rest were soft page events and the same purchases counted several times over by overlapping tags. On one campaign, the platform showed a £7 cost per acquisition while the true figure on fulfilled sales was around £1,500, and that gap is exactly what directed thousands of pounds of spend to the wrong place. Before any decision rests on a number, I take it apart into what was counted, and on what basis.
3. The same number, on every basis
Ask three systems what your CPA is and you will get three answers, because each one counts different things over different windows. Most data arguments inside a business are two people quoting different bases without realising it. In my reports the important metrics appear on each basis side by side, platform-recorded, deduplicated purchases, analytics, fulfilled orders, with a plain statement of which basis the decision should rest on and why.
4. Analyses are designed to rule things out
When performance breaks, everyone arrives with a theory. A good analysis is designed so the data can eliminate suspects. An ads platform change can only move that platform's traffic, so if the fall shows everywhere, the change is cleared. A tag change hits every channel's recorded sales, while a keyword change hits one campaign's, so splitting the data separates the two. I use like-for-like comparison windows, breakpoint scans that find where each series actually broke rather than where everyone remembers it breaking, and control metrics that do not depend on the thing being tested. On one engagement this cleared, in a day, a change the whole business blamed, then located the real breaks weeks earlier where nobody had been looking.
5. Change history is evidence
Ad accounts and tag containers keep records, and I use them. When performance moves, I pull the change history: what changed, when, published by whom, and whether the dates line up with the damage. Findings go onto a dated timeline with a source against every claim, so nothing rests on memory or on any single document. Theories that survive that treatment are worth acting on. Theories that do not get closed in writing, including my own: on one investigation I disproved a theory I had raised myself, and said so.
6. Fixes get monitored so they stay fixed
Every engagement leaves monitoring behind. Scripts check URL health, disapprovals, invalid click rates and zero-sale spend on a schedule. A daily heartbeat flags any day of spend without sales. Alerts are exception-only, so silence means healthy and a message means act. Problems that would have surfaced at month end surface the next morning instead.
7. You see my working
Findings arrive as short written reports, usually two or three pages with charts built from verified aggregates, a stated basis for every figure, and a caveats section that says plainly what the data cannot show. Weekly commentary states the action being taken on each metric, not just a description of the movement. When I get something wrong, the correction goes out with the same visibility as the original claim.
The toolkit
- —Google Ads Scripts and the Google Ads API
- —Google Apps Script and the GA4 Data API
- —Google Tag Manager, consent mode and server-side tagging
- —BigQuery and SQL
- —Looker Studio
- —AI-assisted analysis pipelines for reporting at scale
If this sounds like the opposite of your last agency review, that is rather the point.
Or start with the audit, which is this method applied to your account.