How Vedlora measures AI visibility

Every answer is labeled with its collection mode, provider, model and timestamp, and the two modes are never combined in one metric.

Model API answers come from each vendor's official API with web search enabled: OpenAI's Responses API for ChatGPT, Perplexity's Agent API, Google's Gemini API with Search grounding, Anthropic's Claude API, and xAI's API for Grok.

Consumer-app answers (Google AI Overviews and AI Mode) are what people see in the product, captured by licensed data providers. We never run scrapers against AI products ourselves.

API and consumer answers can differ substantially for the same question, which is why the dashboard asks you to pick one.

Each brand has a name, optional aliases (product names, alternative spellings) and owned domains.

A mention is a whole-word, case-insensitive match of the name or an alias, so “Notion” doesn't match “notional”. Brands whose name is a common word should use a more specific name plus aliases.

Position is the order in which tracked brands first appear. When the answer is a ranked list or table, brands are ranked by the item they appear in, and brands in the same item tie.

Citations are attributed to a brand when the cited URL is on one of its domains, including subdomains. Editing a brand re-analyzes past answers at no charge.

For each brand mentioned, a language model reads the sentence around the first mention and labels how the answer portrays that brand: positive (recommended or praised), negative (criticized or discouraged) or neutral (listed without judgment, or mixed).

Visibility is the share of collected answers that mention a brand. Because AI answers vary from run to run, the same prompt rarely produces the same brand list twice.

We report visibility with a 95% Wilson score interval, which stays accurate for the small sample sizes typical of prompt tracking. Narrow intervals come from more answers: more prompts, more engines, or more samples per prompt.

Share of voice is a brand's mentions divided by mentions of all tracked brands, counting one per answer. Average position is computed only over answers that mention the brand.

Engines sometimes time out, rate-limit or refuse. Temporary failures are retried with exponential backoff; permanent failures are recorded with their reason.

Answers that couldn't be collected are shown as gaps and excluded from metrics, never counted as zero visibility.

When several customers track the same prompt, on the same engine, model, location and period, they share one engine call. Each customer's brands are analyzed separately. This keeps costs down without changing results.