Methodology

Each prompt goes to 4 engines through their official APIs, with web search or grounding enabled so answers are built from live sources:

• · Claude
• · ChatGPT
• · Gemini
• · Perplexity

Each engine is asked with no system prompt and no persona. We are measuring what the engine does with a buyer's question by default, not what it does when we coach it.

We measure through APIs. You are probably wondering about the consumer apps — the actual ChatGPT window, the actual Google AI Overview. Those are not the same surface. They apply extra personalization, different retrieval, memory of your past chats, and A/B tests we can't observe.

So our numbers are directional for those surfaces, not identical to them. We take the API path because it is reproducible, auditable, and stable enough to compare against itself in 90 days. Scraping consumer interfaces produces a number that can't be reproduced next quarter, breaks whenever a UI ships, and violates terms of service.

What carries over well: which brands the engines know in your category, which third-party sources they lean on, and how you rank against your competitors. What carries over poorly: the exact percentage a specific person in a specific session would see.

We read your homepage, infer your category, ICP and likely competitors, then generate buyer-intent prompts balanced across four classes:

• · Category discovery
• · Head-to-head comparison
• · Problem-led search
• · Vendor-specific

Only the vendor-specific class is allowed to name your brand. Most prompts are questions you could plausibly lose — a prompt set you win by construction measures nothing, and it is the easiest way for a tool like this to flatter you into a subscription.

Your prompt set is then frozen and stored. A re-scan replays the same prompts verbatim, which is what makes a before/after comparison real rather than two unrelated samples.

A single answer from a language model is noisy — ask twice, get two answers. The paid audit samples every prompt on every engine three times (40 × 4 × 3 = 480 answers) and reports the spread between passes. If that spread is wide, we say so on your report instead of quoting a false point estimate. The free scan is a single pass over 5 prompts, which is enough to detect a problem and not enough to size it precisely.

Every answer is parsed by a language model under a strict schema to pull out which brands were named, in what order, with what sentiment, and every URL cited. Domain matching — is this citation actually your site? — is done in code, not by the model, because that part has a right answer and shouldn't be guessed at.

Runs that fail outright (a provider outage, a rate limit that survives retries) are recorded as failures and excluded from the denominators. An outage is not a bad score, and your report shows the failed count so you can see the sample shrank.

The headline 0–100 figure is a weighted composite of four measured rates:

• · 40% presence rate — share of answers naming you at all
• · 30% citation rate — share of answers citing your own domain as a source
• · 20% share of voice — your mentions divided by all brand mentions
• · 10% position — first mention scores full marks, decaying to zero by sixth

Presence is weighted heaviest because if you aren't named, nothing else can help you. Citation is close behind despite being rarer, because being the source is what sends traffic and is the hardest thing for a rival to take from you. Every underlying rate is shown on your report, so you can ignore our weighting and read the raw numbers.

• · We do not scrape consumer ChatGPT, Gemini, or Google AI Overviews.
• · We do not claim to know how a specific individual's session will answer.
• · We do not sell placement, and we have no relationship with any engine we measure.
• · We do not reuse your prompt set for anyone else, or publish your results.