A convincing GEO case study should let us trace a result from the original question to the recorded AI answer, the changes made and the business outcome. A screenshot showing a brand recommendation is useful evidence of one appearance. It does not establish sustained visibility or explain what caused it. At Clickspo, we use this distinction to assess generative engine optimisation claims and decide what deserves further investment.
TL;DR: What a useful GEO case study shows
- The same relevant questions measured before and after the work.
- Separate counts for brand mentions, website citations and business enquiries.
- Dated evidence, a clear record of changes and an explanation of limitations.
Start with the customer decision
A Hong Kong business can appear for its own name while remaining absent when customers compare providers. We therefore ask what the campaign was meant to improve. Was it discovery for a service, inclusion in a shortlist or accurate answers about an existing brand?
For example, a hypothetical Central accounting firm might want visibility for questions about choosing an accountant for an SME. Asking an assistant to describe that firm’s brand would test a different task. A study that substitutes easier branded questions halfway through has changed its scope, even if its headline score rises.
We separate commercial discovery, provider comparisons and factual brand questions before interpreting performance. Our guide to auditing competitors in AI search explains how those questions help identify relevant competitors.
Keep the baseline comparable
AI answers can vary across runs. The evidence should record the platform, date, language, question and relevant search setting. English and Traditional Chinese questions deserve their own results because they address different wording and potentially different sources.
| Evidence | What we look for | What it cannot prove alone |
|---|---|---|
| Recorded answer | The actual wording and visible references | Consistent appearances across future runs |
| Repeated prompt set | Comparable questions and collection conditions | How every real customer searches |
| Website analytics | Identifiable visits and recorded actions | Every visit influenced by an AI answer |
| Change log | What changed and when | That one edit caused the entire result |
We also check whether unavailable responses were excluded from the calculation. Removing a difficult question after a failed run can make an otherwise unchanged campaign look stronger. The report should make the denominator clear and compare like with like.
Separate mentions, citations and commercial results
A mention means the answer names the business. A citation means the answer references a source, which could be the business website or another publisher. Neither automatically means the business received a qualified enquiry. We explain that distinction in our brand mentions and source citations guide.
Consider this illustrative example, not a Clickspo client result: a brand appears in 5 of 20 comparable answers before the work and 8 of 20 afterwards. Its observed mention rate moves from 25% to 40%. That is an increase of 15 percentage points, or 60% relative growth. Both descriptions can be mathematically correct, but the report needs the starting counts.
If the prompt set changes to 40 questions in the second period, we would compare the original shared questions separately. Otherwise the apparent movement could reflect the new sample rather than improved performance.
Read the evidence in three layers
Was the business named in relevant answers?
Which sources supported the answer?
Did identifiable visits lead to useful enquiries?
Ask what changed and what else could explain it
A useful case study connects each action to a plausible problem. Correcting inconsistent company details addresses factual ambiguity. Making service pages accessible addresses discovery barriers. Publishing a genuine project example adds evidence that a generic service description lacks.
- Record publication and implementation dates.
- Identify which pages or business details changed.
- Note simultaneous SEO, advertising or publicity activity.
- Allow for changes in AI platforms and source selection.
- Include unchanged or weaker results alongside improvements.
Without a suitable comparison, an improvement after an edit establishes sequence rather than sole causation. We would describe that result as an observed association and keep testing. Our AI visibility and referral measurement guide covers the reporting measures that support this work.
Use the study to choose the next action
If a brand appears but the answer gives the wrong service details, accuracy is the immediate priority. If competitors appear while the brand is absent, we investigate relevant pages, discoverability and the evidence available about the business. If citations increase without enquiries, we examine the question intent and the landing page before commissioning more content.
This is how we would assess a GEO case study: find the customer question, inspect the evidence and connect the next action to the actual gap. Explore our GEO services if you want us to assess those opportunities for your business.
Frequently asked questions
Can we publish a GEO case study without naming the client?
We can anonymise a study, provided the client has agreed to the disclosure and the remaining details do not reveal confidential information. We would retain enough context to explain the market, work and measurement period, while removing customer data, private conversations and commercially sensitive figures. Anonymisation should protect the client without making the evidence impossible to assess.
What if we did not record AI answers before the work started?
We would describe the report as a current benchmark rather than claim an improvement we cannot demonstrate. Existing dated records may help establish a partial baseline, but we would state their gaps. We would then fix the question set and collection conditions for future comparisons.
How many questions should a small business track?
We would start with a manageable set covering its most valuable services, customer locations and buying decisions. There is no single question count that proves a campaign works. A focused set measured consistently is more useful than hundreds of loosely related prompts, and any expansion should preserve the original set for comparison.
Should different AI platforms share one visibility score?
We would keep platform results separate first. If we also create a combined score, we would disclose the weighting and keep it consistent between reporting periods. Otherwise, a change in the proportion of questions collected from each platform could move the headline score without a comparable change in visibility.
