How the audit is run
Everything in a report can be traced to a captured answer or a fetched page. This is the exact procedure.
1. Questions
You give us the questions your buyers ask, or we propose them from your website and you edit them. A good set mixes awareness (“what should I look for in…”), comparison (“X vs Y”, “alternatives to X”), commercial (“best…”, “how much does…”), local (“…in Tallinn”) and one or two with your brand name. The Audit tests 15, the Pro Audit 40.
2. Engines and access
Each question is submitted to ChatGPT, Perplexity, Google AI Overviews, and Claude through the engine’s official API or search surface, logged out, with no system prompt favouring your brand. The only context added is a neutral location hint for your market (“I’m asking from Estonia”). Google AI Overviews are captured from a Google search for the question from your market; when Google shows no AI Overview, that is recorded as “N/A”, which is itself a finding. We never scrape consumer interfaces behind logins and never attempt to get around blocks.
3. Repeat runs and variance
Answer engines are not deterministic. Every question is run 2 times in fresh sessions (3 in the Pro Audit). Both runs are kept; the evidence map shows each one, so a brand that appears in one run and not the other is visible as exactly that.
4. What counts
| Named | The answer text contains your brand name (we match the everyday form you give us plus your domain). |
| Cited | The engine’s returned sources include a page on your domain. |
| Domain only | Your domain appears but the brand name does not, or you are cited without being named. |
| Absent | Neither. |
| Competitor named | A tracked competitor’s name appears in the answer. |
| Accuracy flag | A claim an engine made about you that contradicts the facts you gave us (severity by consequence: prices and regulated claims first), or that we could not verify (“client to confirm”). |
5. Visibility score
60 points for the weighted mention rate (commercial, comparison and local questions weigh 1.25×, brand-name questions 0.75×), 25 for the citation rate, 15 for accuracy minus a penalty per flag. It exists so you can compare this sample with the next one; we do not publish industry averages.
Competitor benchmark and cited sources. Paid reports score every tracked competitor on the same formula, from the same answers: a competitor is named when its name appears in the answer, cited when its own website is among the answer’s sources, and ranked by where it is first mentioned. Competitors are not fact-checked, so they get the full 15 accuracy points; the comparison never flatters you. The report also lists every site the engines cited, typed as your own site, a competitor, editorial, directory/review, forum/UGC or other, and the sites that were cited next to competitors but never next to you. “Cites you” means the site was a source for an answer that named you; we do not read the source pages themselves.
6. Technical readiness
We fetch robots.txt, llms.txt, sitemap.xml, your homepage and up to five pages you choose, as an ordinary crawler would. For each AI crawler we record whether robots.txt allows or blocks it, separating crawlers that feed live answers from those that only feed training:
| Crawler | Owner | Feeds |
|---|---|---|
OAI-SearchBot | OpenAI (ChatGPT search) | live answers |
ChatGPT-User | OpenAI (ChatGPT browsing) | live answers |
GPTBot | OpenAI (training) | model training |
PerplexityBot | Perplexity | live answers |
Perplexity-User | Perplexity (user-triggered) | live answers |
Google-Extended | Google (Gemini training; AI Overviews use Googlebot) | model training |
Googlebot | Google Search + AI Overviews | search and answers |
ClaudeBot | Anthropic (training) | model training |
Claude-SearchBot | Anthropic (Claude search) | live answers |
Claude-User | Anthropic (user-triggered) | live answers |
Bingbot | Microsoft (Copilot + ChatGPT index) | search and answers |
CCBot | Common Crawl | model training |
Applebot-Extended | Apple Intelligence | model training |
meta-externalagent | Meta AI | model training |
Pages are scored for citability: text present without JavaScript, length, headings, an answer-first opening, JSON-LD structured data, FAQ blocks and a visible date. llms.txt is reported but never rated above low impact; its effect is unproven.
7. Fix list and review
Findings are turned into 8–12 fixes ordered brand safety → citation gaps on high-intent questions → technical → content, each with the page to change, the change, effort (S under 1 h, M 1–4 h, L more), impact as a hypothesis, and a suggested owner. A person reads every report against a checklist (counts match the matrix, every flag has a reference, nothing is asserted about a competitor beyond what the captured answers show) before it is delivered.
8. What this is not
Not a ranking. Not a share-of-voice measurement of the market. Not a guarantee of future mentions. Not monitoring (that is a separate product that repeats this method monthly). Not an SEO crawl. Not implementation.
Sources for the numbers on the homepage
- 68% of US Google searches ended without a click, Jan–Apr 2026 — SparkToro / Similarweb
- AI Overviews appear on about 43% of searches — Similarweb, 2026
- 29% of US consumers use chatbots for product research; 41% have bought on a chatbot recommendation — Semrush, July 2026 (n = 2,338)