Fewer case studies, every figure measured.
Most results pages in this category are a wall of percentages with no method behind them. Each study here carries dated figures from named tools: the client’s own analytics, a town-by-town map grid, and AI answers measured on a fixed question set.
Where a figure is still tracking, the study says so instead of dressing a starting position as a result. And the one experiment that went wrong is published below, with its control group.
Free. A short form, then we run the full audit and walk you through it on a call.
These three count what is on this page, and you can check them against the cards below.
We changed one title, tanked a client’s homepage, proved it was us, and put it back.
On 31 July we rewrote the homepage title for Luma Creative, a San Francisco video production company, as part of a title experiment with a control group. By the next morning the homepage’s average Google position had fallen from 16.8 to 34.5, and daily impressions from 889 to 538. The 34 pages we hadn’t touched moved 0.8, which is how we knew the damage was ours and not an algorithm update. On 10 August the old title went back.
The mechanism was visible at the term level: for “corporate video production san francisco” the homepage fell from position 5 to 43 while the corporate page rose from 36 to 15. Google had moved the term to a worse-ranking page.
- measured
- Google Search Console, homepage average position and daily impressions, against 34 untouched control pages.
- window
- Title changed 31 July 2026 · fall measured 1 August · reverted 10 August
- against
- The client's own untouched pages, which moved 0.8 while the retitled set moved 3.1 and the homepage 18.4.
Nobody in this category publishes their misses. We publish ours because a firm that only ever reports wins is telling you about its marketing, and a controlled rollback is better evidence that the measurement works than any win we could show you.
A baseline is a starting position.
Client analytics figures publish with their window and their source, once the client agrees. Our own measurements hold a higher bar: a baseline publishes once, with the question set and window that produced it, and movement gets claimed only after a second capture. Until then the study says tracking, because that’s what’s true.
- instrument
- Client analytics · Caldrin measurements
- what was measured
- Search and website performance from each client's own Search Console and analytics accounts. Map coverage from a query-by-location grid. How often AI names the client, from a fixed question set run across ChatGPT, Gemini and Google AI Overviews.
- how
- Each question runs several times per engine in fresh sessions, with answers stored unedited. Grid positions are checked per town and per service rather than averaged into one number.
- over what window
- Stated on each case page, beside the figure it belongs to. A figure with no published window does not appear as a figure.
- what this cannot tell you
- Being named by an AI is presence, not revenue, and the two are never conflated. Perplexity and Claude are not measured. A client is named only with written permission on file, and a quote appears only with permission and attribution.
Get the same audit these started from.
A full audit of your own business, in two parts: where you actually stand in Google, town by town, and what ChatGPT, Gemini and Google’s AI answers say when a customer asks, scored against competitors you pick. Findings are numbered and ordered by what they cost to fix.
Free. A short form, then we run it and walk you through it on a call. You keep the report either way.