Checking the measurement before the experiment
I asked AI to help measure website improvements. The first task was to check whether our pages were being recorded at all.
On 28 October, I asked AI to help me build a way to improve our website and check whether each change had worked. I wanted to start with cgmkfoods.com: establish a baseline, check the data, then choose the first search improvement and how to measure it.
The first audit found a problem before we reached the writing. Our custom measurement script covered only nine of the seventeen commercial pages we wanted to follow, and the Google Analytics property had not yet received website data. A quiet report could therefore mean that a page was missing from the measurement, rather than that nobody was interested in it. The nine covered pages were a code-level finding, not nine fully verified reporting paths.
I asked the agent to repair that first. It extended the coverage to the agreed seventeen pages and connected Google Analytics. We kept the page list fixed for the checks. The current dated menu also had to be recognised as the same menu page, so changing the week would not create an extra page in that count.
After the change went live, the agent made controlled test visits and checked what reached both analytics services. The checks followed each page through the outgoing request and the account’s own readback. An independent review could trace all seventeen pages in the saved evidence. That gave us a specific result: the agreed measurement paths had passed this technical check.
Those visits were tests, not customers. Opening a text-message draft was still only opening a draft; it did not prove that a message was sent or an order confirmed. We also kept the remaining limits in the record. The reporting filter’s propagation had not been proved, and the seven-day health check and twenty-eight-day baseline had not been completed. More recorded events after repairing coverage could not be counted as business growth.
This changed the order of the work. Before asking whether a new page introduction attracted more people, I needed a record that could support the comparison. AI helped inspect the pages, make the repair and assemble the evidence. The search experiment could then have its own question and observation period, with room for the answer to be that there was not enough evidence yet.
