Methodology

This page explains what our scans actually do, so you can judge how far to trust a result. Where a method has limits, we say so. If anything here is unclear or looks wrong, email [email protected].

• Every scan fetches the page live at the moment you run it. Audit results are never served from a cache.
• We request the page with a Googlebot user agent and retry with a Chrome user agent if the site blocks bots.
• By default we analyse the HTML the server returns. JavaScript rendering (a real headless Chrome) is optional; turn it on for sites that build their content in the browser.
• One page per scan. The crawler follows internal links for up to 30 pages, analysing each page the same way.

• The audit runs 22 modules (on-page SEO, indexability, security headers, accessibility, performance and more). Each module runs pass/fail checks and produces a 0–100 score.
• The overall score is a weighted average of the module scores. Every report has a "How was this calculated?" panel showing each module's share of the score.
• Modules that don't apply to the page are left out of the score rather than counted as failures — for example local SEO on a site with no local-business signals, or PWA checks on a site that isn't an installable app.
• A module that fails to run is also left out and listed as "could not run". If the page itself can't be fetched, the scan fails with a clear error instead of reporting a score.

• Performance is lab data. Speed metrics come from a single test from our server, only when JavaScript rendering is on. They are not real-user (field) data such as Chrome UX Report, and can differ from what your visitors experience.
• Link checks can be blocked. Links that refuse automated requests (401, 403, 405, 429) are not counted as broken, so a genuinely broken link behind bot protection can be missed.
• Security checks see only the page. We inspect headers, cookies and markup. We cannot see server-side protections such as header-based CSRF tokens.
• Heuristics are heuristics. Checks like content depth or keyword use follow common SEO guidance, not Google's actual ranking system, which is not public.

• Free plan: SEOCheckPilot In-house Search, a self-hosted meta-search. It returns web search results, not answers from an AI model.
• Paid plans: Perplexity Sonar (AI answers grounded in live web search) and Tavily (web search built for AI agents), plus three model APIs — OpenAI GPT-4o mini, Google Gemini Flash-Lite and Anthropic Claude Haiku — called without web search.
• API, not the consumer apps. We do not query the ChatGPT, Gemini or Claude apps, Microsoft Copilot, or Google AI Overviews. Model APIs without web search answer from what the model already knows, and published studies find their answers can differ substantially from the consumer apps. Treat those engines as a signal of how well-known your brand is to the model.

• We build questions buyers might ask from your domain and keywords (for example "Best {topic} tools"), and send each question once to each engine on your plan.
• We look for your brand and your declared competitors in each answer, and classify the sources an answer cites.
• Each scan makes at most 60 engine calls. If a scan has more questions than that allows, the extra questions are not sent and are not counted in the score; the scan shows a warning.
• Identical questions sent to the same engine within one hour reuse the same answer, so repeated scans don't trigger new paid calls.
• If an engine is temporarily unavailable, a backup engine may answer. That answer is shown under the backup engine, never under the one you requested.

• Branded questions are excluded. Questions that name your brand ("What is acme?", "acme pricing") mention you by design, so they don't count. The score measures visibility you earn on unbranded questions.
• The score combines citation frequency (40%), citation prominence (20%), topic coverage (15%), share against competitors (15%) and consistency across question types (10%). If there's no competitor data, that factor is dropped and the others re-weighted.
• Each scan's report has a "How was this calculated?" panel with the exact factor values.

• AI answers vary. The same question can get a different answer each time. Research by SparkToro found the same brand list repeats less than 1 time in 100, while how often a brand appears across many runs is much more stable. One scan is a sample, not a measurement — compare trends across scheduled scans rather than reading too much into one result.
• One answer per question per engine. We don't yet repeat questions to average out variation.
• No location or personalisation. Consumer AI products tailor answers to the user; our queries are anonymous and location-neutral.
• Mention detection is text matching. Very short or common-word brand names can be missed or over-counted. Add your exact brand name and aliases in your project's Competitors tab for best results.

Which providers receive your data is listed on the sub-processors page; how we store it is on Security & Data Handling.