Methodology
How we review your website, observe AI answers and turn documented gaps into priorities. Last reviewed: 23 September 2026.
Our AI Visibility Audit has two layers: a review of your site's readiness and a manual test of how four AI assistants describe your brand. We report each layer separately so you can see what was checked, what was observed and what remains unknown.
The readiness score summarizes a review rubric. It is not a percentage of AI visibility, a platform ranking or a measurement of a model's training data. We do not have access to private training datasets or internal ranking systems.
1. Review the technical foundation
The paid audit reviews crawl and index access, structured data, brand links, key-page clarity and buyer-intent content. Its 0–100 readiness score and risk band summarize that review; findings carry a status, confidence level, priority and owner role. Checks that cannot run with the available access are marked not measured.
A high readiness score does not establish that an assistant has retrieved a page or recommended the brand. Those are separate observations from the answer test.
How the free technical scan differs
The free scan uses automated checks on public files and the returned HTML of up to ten sampled pages. It summarizes discovery, crawlability, structure and content readiness, with equal weight for each area. Its text and heading checks are simple review heuristics, not platform eligibility requirements or an assessment of writing quality.
It does not query AI assistants, verify search indexing, inspect every page or replace the paid audit. The scan and the broader audit have different scopes; their scores should not be treated as interchangeable measurements.
llms.txt is optional. The scan records whether the file was found, without awarding points or raising an issue when it is absent. Google says its Search features ignore these files; that statement applies to Google Search, not every AI system. Read Google's guidance.
2. Observe answers under recorded conditions
- Agree on the question set. We use your buyer questions, search data and category context to define the prompts, peers, market, language and location. The question-set size is fixed in the written scope. Branded questions remain separate from questions that do not name your brand.
- Freeze and document the baseline. Exact prompt text, version, assistant, mode, account context and observation date are recorded. Changes to the procedure are disclosed before comparing results.
- Run two manual passes. A consultant tests ChatGPT, Gemini, Perplexity and Copilot in fresh browser sessions, with a new conversation for each prompt. The second pass takes place on a separate day using the same questions and market conditions.
- Keep the full record. Answers are captured verbatim, including brand absences and unfavorable framing. We record no answer, refusal and collection failure separately. If the two passes disagree, both observations remain in the report.
What we record separately
- Brand mention
- The answer explicitly names the brand. Positive, neutral and negative framing are retained; a mention alone is not an endorsement.
- Linked citation
- A source URL shown with the answer. We distinguish your own pages from third-party sources and record which sources appear when your brand is absent.
- Position, competitors and representation
- Where the brand appears, which alternatives are named and how the answer describes it. Factual conflicts are documented against the available evidence.
Results stay separate by assistant, prompt and pass. We do not combine them into a proprietary AI-presence score. Any reported rate includes its count, denominator and observation window.
3. Turn findings into a decision
The audit delivers a baseline report, citation map and prioritized 30-day plan, followed by a readout. Each action names the evidence, owner role, dependency and verification method. Low-confidence recommendations are framed as tests. Waiting or fixing an issue internally can be a valid next step.
Implementation, ongoing monitoring and a later rerun are separate from the diagnostic. Website visits and qualified enquiries also remain separate from answer observations; the audit does not attribute sales to a change in sampled mentions.
What these observations can tell you
They describe a defined question set under recorded conditions. They do not represent every buyer, all personalized answers or your share of all AI searches. Answers can vary by phrasing, product mode, location, account context, model changes and date. Two passes expose variation; they do not make the sample representative of the market.
A citation does not reveal training-data inclusion or prove which source caused an answer. A captured answer is a dated observation, not a promise of future placement. Illustrative examples and methodology walkthroughs are labelled as such and are not client outcome evidence.