1. Define what you want to understand
Imagine your team publishes a guide and wants to understand whether it helps explain its offering in AI search. An isolated screenshot may be interesting, but it cannot tell you what happens with other questions, which sources matter or whether the observation repeats.
Start with a business decision: understanding what information an agency encounters before evaluating a diagnosis, for instance. For KAIRUM, that scope makes a specific need researchable without claiming to represent every marketing conversation on the internet.
2. Choose questions that represent decisions
Cover discovery, comparison and evaluation. Separate questions that name the brand from those that do not: they serve different objectives.
Keep the exact wording and an intent label. If you later change the questions, record that change before comparing periods.
- Discovery: how can I measure whether a brand appears in AI answers?
- Comparison: what should I compare when choosing an AI visibility tool?
- Brand evaluation: what does KAIRUM publish about its diagnostic method?
3. Record where and how you observe
Record the date, language, environment, identified model, search availability and coverage. An API answer and an answer from a consumer interface come from different environments.
Measurements using OpenAI API or OpenRouter must be labeled as API or proxy observations. They are not ChatGPT interface captures. Model names do not establish a commercial integration.
- Question and objective of the set.
- Date, language and geographic context actually available.
- Identified environment and model, without filling gaps by assumption.
- Complete answers, failures and observed coverage.
4. Separate mentions, citations and visits
A brand can appear without its website being cited. A page can be cited without anyone opening the link. A visit needs a separate observation in web analytics.
Record signals separately. Where an answer includes search, distinguish retrieved sources from citations visible in the final text.
| Signal | What it shows | What it does not establish |
|---|---|---|
| Mention | The brand appears in the answer. | That the answer links to its website. |
| Citation | The answer attributes information to a source. | That a person visited the page. |
| Referral visit | A session arrives through an identifiable link. | That it becomes an inquiry or sale. |
Three questions. Three records.
Read the complete answer and the context of the mention.
Keep the URL and check which statement it supports.
Consult authorised analytics and the attribution it actually provides.
5. Read the result within its sample
If you calculate presence, explain which answers count as valid and show the denominator. Errors stay visible as part of coverage.
One run does not represent every possible answer. To read change, repeat compatible conditions and look for patterns before interpreting an isolated difference.
If you changed content and later observed improvement, document both. Their sequence alone does not prove that the content caused it.
6. Turn the reading into a next question
Connect each priority to its evidence. It may mean clarifying a capability, reviewing an outdated source or investigating a question that remains uncovered.
A useful cycle retains its objective, measurement, decision and conditions for another observation. A diagnosis opens a conversation; the team's judgment still matters.
- What we observed and in which answer.
- Which source or content deserves review.
- Which action is proposed and who will review it.
- Which conditions to retain for the next observation.
7. Build a record someone else can review
The difference between an impression and a useful diagnosis is the ability to return to the material. Keep one row per answer, with an identifier connecting the exact question, conditions, complete answer and cited URLs. An executive summary does not replace that evidence archive.
For a first exercise about KAIRUM, choose an agency decision and write four open questions and two directed ones. This is a starting proposal, not a universal sample size. Check whether they cover different decisions before investing in more runs. The question selection article develops this example.
| Field | What to preserve |
|---|---|
| Question | Exact wording, intent and question-set version. |
| Observation | Date, surface, language, answer and execution status. |
| Signals | Mention with context and each citation with its URL. |
| Interpretation | What we observed, what needs checking and the decision it opens. |
8. Review before expanding
Ask another person to take two rows and reconstruct your conclusion. If they cannot find the answer, understand the denominator or tell an example from a result, improve the record before increasing the sample.
Then choose a small action: clarify a definition on a page, better document a product’s scope or investigate an external source. Record what you expect to observe and how you would review it. Google retains SEO fundamentals for its AI features: useful content, technical accessibility and internal links remain part of the work.
The final deliverable can be one page: objective, sample and coverage; two findings with evidence; one limitation that changes the interpretation; and one priority with an owner. A diagnosis is useful when it helps choose the next step and allows a discussion of why.
