How AI decides who to recommend.
- ChatGPT
- Google AI Overviews & AI Mode
- Gemini
- Perplexity
- Claude
What the engines lean on. Roughly in this order.
- 01Search rankings
Most AI answers start with a live web search. Pages that already rank on Google and Bing are what the model reads first.
- 02Google Business Profile
Categories, hours, photos and Q&A feed Google AI Overviews, AI Mode and Gemini directly, and show up in maps results others cite.
- 03Reviews
Volume, recency and what people actually say. Models repeat phrases like “great with kids” when reviews say it often.
- 04Directories and press
Yelp, Tripadvisor, Doctoralia, local guides and “best of” lists are cited constantly. If you're missing there, you're often missing in the answer.
- 05Your own site
Clear service pages, prices, locations and FAQs make it easy for a model to quote you correctly.
Order is a rough guide from what we see in cited sources; weights differ by engine and change over time.
How the MazeRank score works. Same scale on every plan.
Your visibility score (0–100) measures how often, and how high, AI assistants recommend you when people ask the questions your customers ask.
- 1. Each answer earns points. Named first = 1 point, second = 0.85, and so on down to 0.4. Named without a list = 0.5. Not named = 0. Each question's points are averaged across the AI engines that answered it.
- 2. Questions are grouped by intent. Best/top picks (25%), problems (20%), specific services (15%), prices, comparisons, online & nearby, and reviews & trust (10% each). The weights are fixed, so the score always reflects the same balance of searches. If a group doesn't apply to you, its weight is shared among the others.
- 3. More questions sharpen the score, they don't change it. Every plan asks a balanced sample from each group. A bigger plan measures each group more precisely, so the margin of error shrinks.
- 4. Small samples are steadied. A group with only one or two questions is pulled slightly toward your overall average (as if it had two extra average questions), so one lucky or unlucky answer can't swing the score.
Shown as 52 ± 6: we're 90% confident your true visibility is between 46 and 58. Asking more questions (or checking more often) narrows that range; it doesn't lower your score.
A balanced sample: the same mix of intents, fewer of each.
Same expected score. Each group is measured more precisely, so the range is tighter.
Why more questions don't lower the score: extra questions are spread across the same intent groups with fixed weights, so they refine each group's estimate instead of tilting the mix toward harder, long-tail searches. The margin of error shrinks roughly with the square root of the number of questions.
Questions
What is GEO?
Generative Engine Optimization: improving how often and how well AI assistants like ChatGPT, Gemini, Claude and Perplexity mention you when people ask for recommendations. It's the practice, not a trick.
What is AEO?
Answer Engine Optimization. A close cousin of GEO, older term, focused on being the direct answer in featured snippets, voice assistants and now AI answers. In practice the two overlap almost entirely.
What are Google AI Overviews and AI Mode?
AI Overviews are the AI summaries at the top of many Google results. AI Mode is Google's full conversational search. Both draw on Google's index and Business Profiles, so classic Google presence matters a lot here.
How do AI engines pick which businesses to recommend?
They search the web, read the top results, then summarise. That means search rank, your Google Business Profile, reviews, directory listings (Yelp, Doctoralia, Tripadvisor and similar) and your own site all feed in. Each engine weighs them differently and answers vary run to run.
How is this different from SEO?
SEO wins you a link on a results page. GEO is about being named in a single synthesised answer, often with no click at all. Good SEO helps a lot, but the answer also depends on reviews, third-party mentions and how clearly your services are described.
Can anyone guarantee I'll be cited?
No. Nobody controls what a model says, and answers change between runs. Anyone promising guaranteed placement is overselling. What you can do is measure honestly and improve the signals models rely on.
Do llms.txt or schema markup help?
Little proven effect so far. Schema is cheap and fine to have; llms.txt isn't used by the major engines in any documented way. We'd spend your time on reviews, listings and content first.
How does MazeRank measure visibility?
We ask each engine the questions your customers ask, in every place and language you serve, repeat runs to smooth randomness, and score how often and how high you're recommended. We also log which sources were cited and who was named instead of you.
How does MazeRank help improve it?
A fix list ranked by impact: missing directories that engines cite, Google profile gaps, review prompts, and pages your site lacks. The next run shows whether it moved.
See where you stand today.
A free scan asks the AI engines a few of your customers' questions and shows who they name.
Run a free scanPut it into practice
Get recommended and cited in AI answers, and see the visits it brings.
Every client, one dashboard, white-label reports.
Ask ChatGPT, Claude or Cursor about your AI visibility.
REST API and MCP server for your data.
MazeRank vs Profound, Peec AI, Otterly and more.