How LLMVerse measures AI visibility.

When you add a brand, we fetch your homepage and make a single model call that infers your category, your likely competitors, and 12–15 starter prompts — the kind of question a buyer would actually type into an AI assistant.

Most of those prompts are deliberately category-level(“best X for Y”, “top tools for Z”, head-to-head comparisons) rather than prompts that name your brand. A prompt that names you will almost always return you; it tells you nothing about whether AI recommends you on its own.

You approve, edit, or replace every prompt during onboarding, and can add your own any time from the dashboard.

We query ChatGPT (OpenAI), Claude (Anthropic), Perplexity, Grok (xAI), and DeepSeek.

On the engines where live search is enabled, the model browses the web and returns the sources it used — those are the citations you see in your dashboard. The rest answer from their training data, which is still a real signal: it reflects what the model “knows” about your brand without looking anything up.

We do not scrape user accounts, automate browser sessions on logged-in UIs, or use credential-shared workspaces. Every request is a first-party API call billed to our account.

An honest caveat: API answers are not always identical to what you’d see typing the same question into a consumer app. The apps add their own instructions and retrieval on top of the model. Treat these results as a consistent, comparable sample of what the underlying models say — not a pixel-perfect replica of one person’s chat window.

Scans run when you click Run Scanin the dashboard, plus one small scan automatically after onboarding so your dashboard isn’t empty.

Scheduled automatic scans are paused while we move them onto background infrastructure that can run them reliably. Paid plans get them back first, and this page will say so the day they’re on.

Each scan covers a set number of prompts per plan and rotates through your prompt list, so prompts beyond that limit get covered on later runs rather than never.

Each raw answer is passed to Claude Haiku 4.5, which returns a structured record: whether your brand was mentioned, its rank if the answer lists several brands, the sentiment toward it, the competitors named, the sources cited, and the themes discussed.

Positionis only recorded when the answer genuinely ranks or compares multiple brands. A response that only discusses you has no meaningful “position”, so we store none rather than inventing a number.

We keep the complete original answer next to every record. If you disagree with a call, open the response in the dashboard and read exactly what the engine said.

Sentiment is judged on how the answer treats your brand specifically, not the general tone of the text — so a long answer that calls one tool great and another overpriced doesn’t average out into “neutral”.

Each score comes with a one-sentence rationale, stored alongside it, so you can audit any sentiment call rather than trusting a coloured badge.

• Mention rate (40%) — how often you appear at all.
• Position (25%) — how high you rank when answers compare brands, scaled by how often you’re actually ranked. Topping one list out of twenty does not score like topping every list.
• Sentiment (20%) — how favourably you’re described.
• Citations (15%) — how often the answer cites your own domain, not merely how many sources it lists.

Both of those last two rules exist because an earlier version of this score was inflated: any mentioned brand collected near-full position and citation credit. We fixed it, and real scores dropped by around 20 points. We’d rather your number be lower and true.

• One sample per prompt, per scan.AI answers vary between runs of the same question. A single sample is a snapshot, so read trends over several scans rather than one day’s number. Repeat sampling with confidence ranges is on our roadmap.
• Google’s surfaces aren’t covered yet. That means Gemini, AI Overviews and AI Mode. Gemini support is built but switched off; we’ll say so here the day any of them is live.
• No published precision benchmarks.We haven’t built a labelled evaluation set yet, so we don’t quote accuracy figures we can’t back up.

AEO is new. The metrics are not yet standardised. Anyone who tells you they have it all figured out is overselling.

We’re publishing our methodology because (a) you should be able to question any number on your dashboard, and (b) we’d rather get corrected fast than be wrong slowly.

Found something we got wrong, or have a method suggestion? Email hello@llmverse.io.