AI Visibility Research Methodology

This page documents the AI-search visibility methods currently used by SIGNAL Lab and the Netsujo SIGNAL free diagnosis. Where those two paths use different processing granularity, we describe them separately instead of presenting a planned or historical method as the current implementation.

01 / Definition — AIO and AI visibility

In SIGNAL Lab, AIO means AI Optimization (AI search optimization): improving public information so that AI search can understand and cite a company or service accurately. AI visibility covers whether a brand is mentioned, whether its official site is cited, and whether the business is described accurately.

02 / Engines measured

• SIGNAL Lab periodic observation — OpenAI Web Search (gpt-4o); answer text and cited URLs are recorded.
• Current free-diagnosis question observation — OpenAI Web Search (gpt-4o-mini). Comparison-axis extraction uses claude-haiku-4-5 as a separate internal analysis step, not as the AI-search observation itself.
• Perplexity Sonar — an adapter exists, but it is not enabled in the current public free-diagnosis path. The provider/model/search/locale/region/executedAt values saved for each run are authoritative for that run.

These are API-based observations and are not guaranteed to reproduce a consumer AI product UI exactly.

03 / Question set and execution count

The published SIGNAL Lab baseline is a historical fixed-point observation using 20 fixed questions across four categories: company recommendation, problem solving, purchase decision, and target-company fit.

The current free diagnosis uses Opportunity Trace V2 and observes the questions shown to the customer on the confirmation screen. Each question normally runs once. A network error, rate limit, or temporary provider failure may be retried at most once after the cost-budget gate; that retry is recovery from a transient failure, not an independent repeat sample.

If a comparison target cannot be established, one broader category observation may be added. That broadening is also bounded to one additional observation. The questions and execution conditions actually used are retained with the report.

04 / Free Diagnosis V2 judgement criteria

• Brand mention — observe whether the target company or service appears in the answer text and reference information.
• Official-site citation — determine whether cited URLs include the target official domain. Mention and citation are separate observations.
• Information accuracy — only observations in which the company appears are judged. The sentence that mentions the company is checked against fetched official-page text. A contradicted judgement requires verbatim evidence that actually exists in the official-page text; if that evidence cannot be established, the result remains unverifiable rather than being called misinformation.
• Comparison candidates — start from references actually cited in AI answers, remove the target domain and domains classified as non-competitors, then prefer domains appearing across multiple questions and keep at most three candidates.

05 / SIGNAL Lab v3 stage-based observation

SIGNAL Lab v3 prioritises the stage at which visibility breaks down rather than treating one total score as the primary answer. A reference aggregate can be shown as a secondary indicator, but unmeasured stages are excluded rather than silently counted as zero.

06 / SIGNAL Lab v3 separates citation selection and absorption

Citation selection

Whether retrieved information is actually presented as a citation in the answer.

Citation absorption

Whether a presented citation is actually reflected in the claims made by the answer.

07 / Accuracy granularity differs by execution path

The SIGNAL Lab v3 methodology model decomposes an AI answer into atomic, verifiable claims and evaluates supported claims against all verifiable claims. If there are no verifiable claims, accuracy is not inferred from a proxy.

The current Free Diagnosis V2 implementation works at a coarser sentence level: sentences that mention the company are checked against official-page text and aggregated as supported, contradicted, or unverifiable. A contradicted result requires verbatim official-page evidence. We do not describe this sentence-level Free Diagnosis V2 judgement as the same atomic-claim precision used by the SIGNAL Lab v3 methodology model.

08 / Evidence strength and source type

Research and experiment claims distinguish established evidence, observed-only findings, hypotheses, and refuted claims. Source type is also recorded so that a platform or primary source is not presented as equivalent to a Netsujo hypothesis or self-observation.

09 / Handling variability

• AI answers vary by run and over time, so a single observed answer is not treated as a universal result.
• Each free-diagnosis question normally has one observation. A transient-error retry is bounded to at most one and is not counted as statistical repeat sampling.
• Fixed-point SIGNAL Lab comparisons keep the question set, engine, conditions, and methodology version explicit.
• A change in an observed answer after a web change is not, by itself, treated as proof of causation.

10 / Limits of this research

• Published v1 fixed-point data includes historical one-run-per-question observations; figures from materially different methodology versions are not directly compared.
• The current free diagnosis is not a repeated-sampling experiment: normal execution is one observation per confirmed question, with only bounded recovery from transient failures.
• API observation does not fully reproduce consumer-product login state, history, or personalisation.
• Indexing and AI-answer update lag cannot be controlled by Netsujo.
• Unavailable or unverifiable information remains unmeasured or unverifiable; it is not converted to zero or “no problem”.

11 / Methodology versions and history

When the question design, judgement criteria, execution conditions, or engine changes materially, we record the version and do not silently compare unlike conditions.

12 / Disclaimer

This methodology and its observations do not guarantee search ranking, inclusion in an AI answer, citation, recommendation, or enquiry volume. Each observation is a result under the conditions recorded at that point in time.