Most case studies in this field give a percentage and a logo. They do not say which questions were tested, how many times, on which engines, or what the starting point was. A reader cannot check any of it, so the number carries no information.
This page shows the approved case studies first, then explains how we measure, what each result can and cannot prove, and how to read anyone's case study, ours included, before you rely on it.
| Business | The question asked | Where | What came back |
|---|---|---|---|
| Reset with Ola | burnout coach geneva | Website second in the results, and the first name in the AI Overview | |
| Reset with Ola | An English speaking burnout coach in Geneva | Google AI Mode and Gemini | Named first on both |
| Physio Buddy | Best physiotherapist in Chandkheda for lower back pain | ChatGPT | Named first |
| Physio Buddy | Best physiotherapist near Chandkheda | Google AI Mode and Gemini | Named second on both |
| Bolo Media | A saree shop in Surat wants Gujarati Reels and ads | ChatGPT | Named first |
| Bolo Media | Gujarati product video ads with an AI voice | Google AI Mode | Named first |
AI answers change between runs and between countries. The detailed case study shows the misses as well as the wins.
Request a full case study by email
01How we measure AI visibility
Every result we report rests on the same four parts, so you can repeat the test yourself and compare.
Named prompts
We write the buyer questions with the client and keep the list fixed. A prompt is never swapped out after the first run because it gave a bad result.
Three engines
Each prompt runs on ChatGPT, Perplexity and Gemini, because the same brand can appear on one and be absent on another.
Repeated runs
Each prompt runs more than once, in a fresh session. Answers change between sessions, so one run is an anecdote and several runs form a range.
Screenshots
We keep the answer as it appeared, with the sources it listed, before any change and after.
From those runs we record three counts: how often the brand is named, how often a page on the client's site is cited, and how often the description is correct. We keep the counts separate, because a brand can improve on one and not the others.
| Measure | Definition we use | Why it is separate |
|---|---|---|
| Mention rate | Share of test runs where the answer names the brand | Shows whether you are part of the answer at all |
| Citation rate | Share of test runs where the answer links a page on your site | Shows whether your pages are used as sources |
| Accuracy | Share of mentions where price, product and location are stated correctly | A wrong mention can do more harm than no mention at all |
| Consistency | How much the above change across repeat runs | Tells you whether a gain is stable or noise |
Metric set adapted from Scalenut, AI visibility tracking success metrics.
02What a case study can and cannot prove
A case study can show that a specific set of changes was followed by a specific change in answers for one brand, in one category, over one period. That is useful evidence of a method working. It cannot show that the method will work for you, and it cannot rule out other causes.
It helps to separate three kinds of evidence. A case study is evidence about one brand. A large study is evidence about a pattern: Seer Interactive, tracking 53 brands, found that being cited in an AI Overview earned 120% more organic clicks per impression than not being cited, and Ahrefs found branded web mentions correlate 0.664 with AI Overview visibility across 75,000 brands. A test of your own questions is evidence about you. The last one is the only one that tells you what to do.
There is also a base-size problem. A jump from one citation to four is a 300% rise and means almost nothing. A rise from 8 named runs in 100 to 31 in 100 is a smaller percentage and means much more. This is why every figure we report comes with the number of runs it is built on, and why we distrust any percentage that arrives without one.
03How to read any AI visibility case study
In the AI SEO case studies we reviewed, the same gaps came up repeatedly: no baseline, no time frame, no sample size, and metrics such as "citation" or "engaged session" used without a definition. One case card we read quoted the same result as two different percentages. Use the checklist below on ours and on anyone else's.
Eight questions to ask of any case study
0 of 8 done04Why we do not quote a single percentage
Google says no one can guarantee a number one ranking, and AI answers vary between sessions even for identical questions. A single percentage hides that variation. We report a range from repeated runs, and we compare it with the baseline range, because a gain that sits inside the baseline's own variation is not a gain.
Correlation is a second limit. Ahrefs itself notes that correlation is not causation in its brand mention study. When a brand's mentions rise after we fix its facts and sources, that is consistent with the work causing it, and we say so with that wording. We do not claim proof.
What case studies usually give you
- One percentage and a logo
- No list of questions tested
- A time frame described as "a few months"
- A single run, or no statement of how many
What we show
- The question list, so you can run it
- The engines and the number of runs per question
- Before and after answers with the sources listed
- A range, with the limits stated beside it
05How each of our case studies is structured
Every case we publish follows one order, so you can compare one with another and check each against the checklist above.
Starting position
The questions, the engines, the number of runs, and the baseline range for mentions, citations and accuracy.
Changes made
What we fixed on the client's site and which outside sources we worked on, listed plainly.
Re-test
The same questions, engines and run count, with the new range beside the old one.
Limits
What else changed in the period, what did not move, and what we would not conclude from the result.
Scroll sideways to follow the stages.
06What we will not do with a case study
We do not invent a result, round up a range into a single figure, or present a gain as proof that the same work will repeat. We do not show a result without the questions behind it. We also do not use a screenshot of one lucky run as the headline, because a run that went well tells you less than ten runs that show a spread.
If a result is mixed, for example mentions rose but accuracy did not, the report says so. A case study that only reports gains is a sales page. The useful ones also say what the work left undone, so you know what to ask for next.
07What to do with this
If you want to know where you stand today, the free AI visibility checker gives a first look and the AI visibility audit runs the full method on your questions. The measurement guide explains each count in more detail, and the AI search statistics page collects the sourced studies. For what the work costs, see the pricing page; for the services themselves, see generative engine optimization services. If your problem is that ChatGPT does not mention you at all, start with brand not showing in ChatGPT. The overview of the whole topic is on the AI visibility hub.
