Methodology: evidence with boundaries

1. Select and version the questions

Start with the business decision, category and intended audience. Use customer research or first-party query evidence when provided with permission. Otherwise label questions as analyst-authored, representative questions—not measured search demand.

Keep branded questions (which name the business) separate from unbranded discovery questions and informational citation opportunities. Freeze a versioned prompt set before the benchmark; record changes rather than silently editing difficult questions.

2. Record the test conditions

For each response, record the timestamp, product surface, consumer UI or API, model/configuration when disclosed, search mode, language, region/settings where available, prompt version and repetition. Use “unknown” for settings the product does not reveal.

Free and paid consumer tiers require actual consumer-surface tests under comparable conditions. Different API models are not a substitute. Separate search-enabled and non-search responses; do not pool them into one comparable rate.

Repeat prompts in clean sessions using a repetition count agreed before testing. Record prior conversational context if a clean session is impossible. Keep the full response and source URLs, including failed runs and partial responses.

3. Classify observable evidence

Mention

An explicit business name or reviewed alias in the answer. Ambiguous matches require a review note.

Recommendation

An explicit suggestion that the business suits the request. A neutral list or passing mention is not automatically a recommendation.

Citation

A source URL presented by the product. Record the cited domain and relevant claim; a citation can support criticism or a factual comparison without endorsement.

Factual accuracy

A claim checked against identified, dated evidence. Mark unsupported or unverifiable claims as unknown instead of assigning them a passing score.

4. Make the denominator visible

Report eligible successful responses with a signal divided by all eligible successful responses in the same stratum. Also show attempted runs, failures, exclusions and the exclusion reasons. Count a response once per signal even if the brand appears repeatedly; retain raw occurrences separately if useful.

For example, “3 of 12 successful unbranded responses recommended the business; 2 additional attempts failed” is an illustrative format, not a Takusen result. Rates based on few runs are unstable. Do not treat missing data as zero or hide failures in a headline percentage.

5. Interpret without overclaiming

One panel measures only its selected questions and conditions. Mention order is not automatically preference or rank. Sampling and model changes can explain differences between runs. Before/after movement alone does not establish causation; preserve comparable conditions and document competing explanations.

Keep synthetic tests separate from first-party visits, inquiries and sales. Those sources answer different questions and do not identify every AI-mediated interaction.

Taku Score status

No implemented, documented universal Taku Score was found in this website repository. We do not promise a 0–100 visibility score. Separate counts and evidence are more interpretable until a scoring specification and validation exist.

Use the measurement design guide or read the technical sample report.