How AI models find answers
An AI answer can draw on information learned during training, sources retrieved for a request, or tools made available in that product. These are different paths to an answer. A brand appearing in one source does not ensure it will appear in the answer.
Training and retrieval
Training shapes a model before a user asks a question. Publishing a new page does not insert it into a model's learned knowledge or establish when a future model will use it.
Retrieval supplies information for a particular request. A system may search or read sources, but the exact behavior depends on the product, mode, prompt, and available tools. Indexing makes a page available to a search system; it does not prove that an AI answer selected it.
Avoid assigning a universal update schedule to ChatGPT, Perplexity, Gemini, or Claude. Check the actual product mode and the sources shown with the answer being tested.
Follow the source evidence
When a sampled answer includes citations, open the referenced URLs. Check that the source exists, that it says what the answer claims, and that it actually refers to the intended brand.
Useful questions include:
- Is the answer citing your website, an independent article, a review, or a discussion?
- Does that source contain current and accurate information?
- Is a competitor mentioned in the source or only in the generated answer?
- Did the answer recommend the brand, mention it neutrally, or criticize it?
Some answers expose little or no source evidence. Record that limitation rather than inventing an explanation for why the model mentioned a brand.
Why public discussions are worth investigating
Relevant discussions can contain useful questions, comparisons, and firsthand accounts. They are possible sources to inspect, not a shortcut to an assured AI recommendation.
Follow community rules, disclose relevant affiliations, and never invent personal experience. A contribution should help readers even if no AI system ever retrieves it.
What share of model means
Share of Model (SOM) measures brand mentions or recommendations in the successfully completed, relevant samples being tested. For a simple example, eight mentioning answers out of forty comparable successful answers is 20%.
Record the prompts, engines, market, evaluation rules, and successful sample count. Failed or skipped checks are coverage information, not successful answers with no mention. Inspect the actual answers alongside the summary rate.
There is no universal starting range, category rank, or assured 90-day increase. A low sampled rate does not prove that a model knows nothing about the brand.
In Outrigger
Use Sonar to review the available sampled answers, sources, and coverage. Keep scope consistent across follow-up runs and record changes to the prompts or brand setup. The six Academy topics help organize work; the current audit excludes Press from scoring.