Short answer: We ask Google Gemini with live web search turned on, the questions real customers ask about your type of business in your city. We record the answers and the sources behind them. Then we score your business on four weighted dimensions: Visibility 30%, Position 30%, Sentiment 20%, and Sources 20%. These combine into an AI Visibility Score from 0 to 100 and a letter grade from A to D. The result is a snapshot from a single test, not a guarantee.
This page explains what we measure, how we measure it, and where the method has limits.
Which AI engine do we use?
All checks and reports currently run on Google Gemini through the Gemini API, with Grounding with Google Search enabled. With grounding, the model can run live Google searches and return citations for the web pages it used.
We don’t currently test ChatGPT, Perplexity, Copilot, Claude, or Google AI Overviews directly. Your customers may use any of them. In BrightLocal’s 2026 survey, ChatGPT was the AI tool consumers used most for local recommendations (31%), followed by Google’s AI Mode (23%) (BrightLocal). Each assistant has its own model, search provider, and ranking logic, so a Gemini result tells you how one major AI system sees your business at one point in time. It’s a useful, consistent reading. It isn’t the full picture.
What questions do we ask?
We write questions the way a local customer would ask them, for example: “Who is the best dentist near me for a nervous patient?” or “Which plumber in [city] can come out today?”
- Free check: 8 customer-intent questions.
- Full report: up to 16 questions across 7 intent clusters:
| Intent cluster | What it represents | Example of the kind of question |
|---|---|---|
| Discovery | First-time searching for a provider | “Best [category] in [city]” |
| Comparison & Value | Weighing options against each other | “Which [category] in [city] is worth the price?” |
| Trust, Reputation & Reviews | Checking reliability | “Most trusted [category] in [city] with good reviews” |
| Price & Insurance | Cost and coverage questions | “Affordable [category] in [city] that takes my insurance” |
| Availability & Booking | Urgency and convenience | “[Category] in [city] with same-day appointments” |
| Services & Specialties | Specific needs | “[Category] in [city] that specializes in [service]” |
| Brand Probes | Questions about your business by name | “What do people say about [your business]?” |
The exact wording is adapted to your category and city. The full report shows up to 16 questions, each with the AI’s verbatim answer and the sources it cited.
How is the data collected? (Two-stage collection)
We separate finding out from grading.
Stage 1: Research (grounded). For each question, the model is told to check current information with live Google Search and then write the natural answer it would give a customer. We save the answer text and the source URLs the model cited.
Stage 2: Scoring (evidence-only). Your business, and every competitor that came up in the research, is scored question by question. The scores cover mention, recommendation strength, position prominence, sentiment, and factual accuracy. This stage uses only the evidence collected in Stage 1 and runs no second live search. That keeps the grading tied to what the AI actually said and stops the model from inventing new sources while it scores.
In a full report, competitors are scored on the same questions you are. This is how the side-by-side comparison of 2–3 same-city, same-category competitors is built.
How is the score calculated?
Your AI Visibility Score (0–100) is a weighted total of four dimensions:
| Dimension | Weight | What it measures | Main signals |
|---|---|---|---|
| Visibility | 30% | Your share of AI conversations: how often you’re mentioned or recommended across customer-intent questions | Mention rate, recommendation rate |
| Position | 30% | How prominently you appear when the AI does recommend you. If you aren’t mentioned, this scores zero | Position prominence in the answer |
| Sentiment | 20% | Whether the AI’s description makes customers want to choose you, and whether the facts are right | Description tone, information accuracy |
| Sources | 20% | The “resume” the AI reads before recommending you: how strong and consistent your web footprint is | NAP (name, address, phone) consistency, website content, Google Business Profile completeness, directory coverage |
A mention counts only when the AI actually names your business or clearly refers to it. A business can rank #1 on Google Maps and still be missing from AI answers if the sources about it are thin, inconsistent, or hard for a model to cite.
What do the grades mean?
| Grade | Score | What it usually means |
|---|---|---|
| A | 85–100 | The AI mentions you often, describes you favorably and accurately, and backs this with strong sources. Work on keeping it that way. |
| B | 70–84 | You show up for a good share of questions, with clear gaps in some intent clusters or dimensions. |
| C | 55–69 | You appear sometimes, but competitors are named more often or more prominently. |
| D | Below 55 | The AI rarely names you, names you without recommending you, or relies on weak sources. Usually there’s a lot of room to improve. |
A low grade doesn’t measure the quality of your service. It measures how visible and citable your business is to an AI system at the time of the test. A low grade is often the most useful result, because it shows exactly where the gaps are.
What happens when live search doesn’t run? (Reduced-confidence handling)
The Gemini model decides on its own whether to run a Google search for a given prompt (Google’s documentation describes this step). Sometimes it doesn’t search, even when we ask it to. An answer written without live search reflects the model’s background knowledge, not the current web, so we treat it as lower confidence.
Here’s how we handle it:
- Automatic retry (free check). If the live-search step doesn’t trigger on a free check, we retry automatically and then re-run the check.
- Clear labeling. If live search still doesn’t run after the re-run, the free-check result is marked reduced confidence / preliminary. It isn’t presented as a complete scorecard.
- No sale on a failed free-check result. We don’t offer the paid report for a reduced-confidence free-check result. You’ll be invited to run the free check again later instead.
- Paid reports: retry or replace per question. Inside a paid report, if a single question fails to trigger live search, returns an answer that is too short, or is empty, we automatically retry that question. If it still fails, we replace it with another question from the same intent cluster. Each question gets at most 3 attempts (original plus retries/replacements). The report is only delivered once the full set is complete.
- If a paid report still can’t be delivered. The rare case where a report can’t be completed falls under our existing refund policy for undelivered reports (full refund). There is no separate rule.
What’s in the full report?
- Executive summary
- Methodology (this protocol, as applied to your business)
- Question-by-question evidence appendix: up to 16 questions, each with the AI’s verbatim answer and cited sources
- Competitor deep dive (2–3 same-city competitors scored on the same questions)
- Gap analysis: missed questions, weakest dimension, weakest intent cluster
- Prioritized 90-day action plan
- Ready-to-use sample content: Google Business Profile description, FAQ Q&A, structured-data suggestions
- Re-measurement plan
- Collection appendix
What are the limitations?
We’d rather you know these before you rely on a score:
- It’s a snapshot. Results reflect model API responses at the time of testing. AI answers change as the search index, the model version, and your own online footprint change. If you run the same check again, you may get a different result.
- API, not consumer app. We query the developer API. Consumer apps like the Gemini app, ChatGPT, or Google AI Overviews may use different models, settings, personalization, account history, and location signals.
- One engine. We currently test Gemini only. Other assistants may recommend different businesses.
- Location is simulated through the question. We name your city in the question. We don’t simulate a phone sitting in a particular neighborhood.
- Small question sets. Up to sixteen questions in the full report (eight in the free check) is a sample of how customers might ask. It’s not every possible phrasing.
- Scoring uses a model. The evidence-only scoring stage is designed to be consistent and traceable, but it’s still automated judgment. That’s why we show the verbatim answers and sources, so you can check every call yourself.
- No guarantees. Following the action plan can improve the signals AI systems rely on. No one can guarantee that any AI system will mention or recommend your business.
How should you re-measure?
Because single runs vary, look at direction over time, not one number. A sensible rhythm is to re-check after you’ve finished a meaningful set of improvements (for example, after 60–90 days) and compare dimensions and intent clusters, not just the headline score. Each full report includes a re-measurement plan.
This methodology describes our scoring model as of October 10, 2026. We’ll update this page and its date whenever the method changes.
Results are a snapshot of model API responses at the time of testing and may differ from consumer-facing AI products due to model versions, personalization, location, and timing.