How Beacon measures AI answers

What Beacon asks, and where

You track questions for one business website. General questions do not name you. They show what AI tells someone who has not asked about your business by name. Questions that name you help measure how AI describes you, and are counted separately.

The AI search view covers ChatGPT, Google AI Overviews, Google AI Mode, Gemini and Copilot in Bing. Beacon also asks Perplexity, Claude and Microsoft Copilot (in pilot) directly, with web search, and shows those answers on their own line, outside the AI search grade. Checks run for the United States, in English; you can track up to 5 regions per business website and compare them in All regions.

AI search and AI models are separate views. Their answers and grades are never added together. Each figure identifies the questions, AI services and dates it counts.

What each number means

Named is the share of general answers that name your business. Share of voice is your share of recorded mentions of you and the competitors you track, across general answers. Cited is the share of answers that cite your site's pages as sources. Naming your business and citing your site are counted separately. Named first is the share that name you before any competitor you track. Sentiment is how AI describes your business in answers that name you. Website readiness checks whether AI crawlers can reach and read your site. It does not show whether an AI service used your pages.

The AI visibility grade combines measurements from general answers with website readiness into a score from 0 to 100. Its rules set the weights and handle missing measurements. Beacon shows a grade only when it can measure at least 60% of its weight. Below that, it reads "Not measured". Questions that name you do not add points. The grade without sentiment is beacon-visibility-grade-v1. Sentiment counts only while its quality check passes, using grade version beacon-visibility-grade-v2. Beacon keeps comparisons within the same grade version. Sentiment uses beacon-sentiment-v2; report and email figures use beacon-brief-derived-v6.

How sure Beacon is

Each share comes with a likely range at a 95% confidence level. A wide range means the answers leave more uncertainty about the share.

Beacon shows a percentage only with at least 30 answers and a narrow enough range. Otherwise it shows the count, such as "5 of 18", without a rating. When no answers name you, or all of them do, the range still allows for uncertainty.

A forecast estimates future checks from earlier answers. Its shaded band covers 80% of the model's predicted results, up to four checks ahead. An isolated unusual check has less influence than several unusual checks in a row. The forecast assumes conditions remain similar; it is an estimate, not a promise.

To assess a change, Beacon compares the same questions across checks. It tests whether the difference is larger than the variation expected from those answers, and accounts for checking several measures at once. A change after an action does not prove the action caused it.

The grade has a likely range too, calculated by repeating the calculation with different selections of your questions. A grade of 90 or above reads "Good"; from 50 up, "Needs work"; below that, "Weak". When its range crosses a boundary between ratings, Beacon says so.

The same numbers everywhere

Each check asks your saved questions on their daily or weekly schedule and keeps the full answers. Once a check finishes, its saved answers support the dashboard, PDF and report email. Compare the same check and view when reading their numbers.

Failed or out-of-date answers do not count as answers that failed to name you. Each figure shows how many answers it uses. A measure needs at least 30 answers, with a narrow enough likely range, before Beacon shows a percentage.

Automated tests check that identical answers do not produce unexpected change alerts, and that saved report figures match the answers behind them. Each week, a separate check compares sentiment labels with labels from three AI reviewers. Agreement between AI reviewers does not establish that a label is correct. Sentiment counts toward the grade, and the Sentiment page shows its headline, only while the latest check passes, covers at least 100 answers and is no more than 15 days old. If the labelling model changes, sentiment leaves the grade until the new version passes a check.

What Beacon doesn’t measure

Beacon measures answers to the questions you track. It does not measure how often people ask them, or every answer a customer might see. Results apply to the AI services, location, language and dates shown. Personal settings and AI-service changes can produce different answers. A higher grade does not guarantee more visits or sales. Use your own traffic and conversion data to assess those results.

Where Beacon uses AI

The AI services being measured supply the answers. Separately, an AI model reads answers that name you and labels how they describe you. It uses the answer's own words as evidence; it does not write quotations. Its version appears with the sentiment figures, so you can identify changes in how answers were labelled. AI labels can be wrong; read the saved answer when a label affects a decision.

Versions

Each measure has a version. When its calculation changes, its version changes too. These dates show when each version took effect in the code.

Questions about the method

Write to beacon@prompeteer.ai with a question about the method, or to have any figure traced back to the answers it came from.