How to Judge AI Recommendation Case Studies Before You Trust Them

A trustworthy case study about AI recommendations shows the exact questions tested, which assistants were checked, the answers before and after, what changed in between, and how long it took. If a case study only shows a dramatic percentage with no prompts, dates or screenshots, treat it as marketing, not evidence.

Why AI visibility case studies are hard to get right

AI answers vary. The same question can give different results on different days, in different accounts, or with small changes in wording. Models are updated, and assistants that search the web see a changing internet. So a before-and-after comparison needs care. One lucky screenshot proves very little.

Checklist: what solid evidence looks like

• Named prompts: the exact questions, such as “best vegan restaurant in Lyon for groups”, not just “relevant queries”.
• Named assistants: which of ChatGPT, Gemini, Claude, Perplexity, Copilot or Google AI Overviews were checked.
• Repeated checks: each prompt tested several times, with results summarised (for example, “mentioned in 6 of 10 runs”).
• Dates: when the baseline and follow-up were taken.
• What was done: specific changes, such as new pages, corrected listings or press coverage.
• Honest limits: what did not work, and other factors that may have played a role.

Red flags

• Large percentages with no base number (“+400% AI visibility”).
• No mention of which prompts or assistants were used.
• Claims that the provider has special access to AI companies or can place clients in training data.
• Results described as guaranteed or repeatable for everyone.
• Anonymous clients with no detail about industry, size or market, combined with very strong claims.

Questions to ask a provider about their case studies

1. Can I see the full prompt list and raw answers?
2. How many times was each prompt tested?
3. Were there other changes at the same time, such as a rebrand, new funding or big PR?
4. How similar is that client to my business?
5. What happened in cases where the approach did not work?

A provider who answers these openly is more credible than one with the most impressive headline.

Run your own small case study

The best evidence is your own. It does not need to be complex:

1. Choose ten buyer questions and three assistants.
2. Run each question three times and record whether you appear and how you are described.
3. Make one or two clear changes, such as publishing an answer page or fixing inconsistent listings.
4. Repeat the same checks after four to eight weeks.
5. Write down what changed and what else happened in that period.

Our guide to measuring AI visibility and brand mentions explains how to record results, and the ROI of AI search optimization covers how to link visibility to business results.

Reading results fairly

When you look at your own before-and-after numbers, be careful not to over-read them. A small rise from two mentions to four in a few weeks might be real progress, or it might be normal variation. Look for changes that hold across several checks and several assistants. Also look at quality, not only counts: are you now described more accurately, or cited as a source for a specific question? Those changes are often more meaningful than a single number.

Why we focus on methods, not success stories

Every business starts from a different place, and no one controls AI answers. That is why Keboost shows you your own data rather than promising results based on someone else’s. A good method, applied consistently, is a better guide than a headline.

How Keboost helps

Keboost records your prompts across ChatGPT, Gemini, Claude, Perplexity and Copilot over time, so you can build your own honest before-and-after view as you add content. Run a free AI visibility check to capture your baseline today.