The methods, in full.

A before and after on the changed pages credits the change with everything else that happened in the window: seasonality, a core update, a competitor going offline, a good month for the industry. On test data, an injected 90% lift during a 52% sitewide decline reads as minus 13.7% before and after, and plus 88% against controls. Only one of those is the effect of the change.

So each changed page is matched to untouched pages on the same property with similar traffic before the change, within a factor of two, up to five controls each. Matching is without replacement: one control page backs exactly one changed page, so a single quiet page cannot underwrite several claims.

The controls give the expected result: what the changed page would have done if nothing had changed. The measured effect is the gap between what happened and that expectation, as clicks per day over a stated window.

Then a placebo test. Sets of untouched pages with the same traffic profile are treated as if they had changed and run through the same estimator, a thousand times. The spread of those results is what the property does with no change at all. A result is only called when the real measurement sits outside that spread.

Three answers are legitimate: it worked, and here is the number; it did nothing this comparison can detect; or there is not enough traffic yet, and roughly how much longer it needs.

One answer from one assistant proves nothing. Answers vary run to run, so a screenshot is an anecdote.

AIO Copilot tracks a set of questions a buyer would ask, and asks each engine each question many times: on the demo client, 21 questions, 10 times per engine, once a month, through each model's API with web search. The share of answers that name the client is reported with a 95% interval, beside the rival each engine names most.

Sampled figures are labelled sampled. Measured AI figures are shown separately and are the only AI figures a report may call measured: impressions in AI Overviews and AI Mode from Search Console, citations in Bing's AI answers from Bing Webmaster Tools, sessions from assistants in GA4, and AI crawler requests from the access log.

The app samples models through their APIs. It does not read the consumer ChatGPT app, and it says so wherever the figure appears.

Spend on a search term with no recorded conversion is measured spend in the window. It is not a saving: the click you stop buying may have been the one that converted later. Negatives are proposed for review, never applied automatically.

A negative's match type decides what it blocks. An exact negative blocks only that term; a phrase negative blocks any term containing that run of words; a broad negative blocks any term containing all its words. Every proposed negative shows the other terms it would also block.

When a search the account pays for also ranks organically, that is shown as a test to run, with the organic position beside the spend. Pausing the ad does not hand over the organic click.

Zero conversions across a whole account almost always means conversion tracking is broken, not that every term failed. That is checked first.

A dollar figure is clicks times a value per click, and the working is shown: $3,514 a month is 67.9 more clicks a month, measured, times $51.75 a click, estimated. The measured part and the estimated part are labelled.

Where the client keeps an intake list, leads and signed work are read from it, so the report's money figure is the client's own.

An estimate never enters a total with a measured figure, and nothing is annualised or forecast.

Search Console tables stop at 1,000 rows. A page missing from a capped table has unknown traffic, not zero, and is excluded and counted rather than scored as a collapse.

Page level impressions overstate the property total, because one results page that shows several of your pages counts once per page. Property totals come from the device or date tables instead.

Google Ads exports carry preamble lines, trailing total rows and dashes for missing values. A dash is never read as zero.

An export with a query or country filter is a slice, not the property, and is labelled as one.