Method

How we measure, and what
the numbers don’t tell you

Don’t trust our numbers. Check them. Every score opens the answers behind it, with the date, the engine and the prompt. This page tells you how each one was produced, where it’s weak, and what we can’t see at all.

Book a demo Back to AI Visibility

The LLMagnet measurement methodVersion 1.0 · last updated Nov 1, 2026

v1.0

01What actually runs
02How many runs stand behind a number
03The same question, twice
04What each word means
05What we don’t measure
06What actually works
07Sources and changes

What actually runs

Your prompts, every engine, every region you track. One scan, read the same way each time.

The prompts: The questions you track: written by you, suggested per buying stage, discovered automatically, or taken from your own Search Console demand. Nothing is scanned until you approve it.
The engines: Seven: ChatGPT, Claude, Gemini, Google AI Overviews, Perplexity, Copilot and Grok. Starter covers three. Copilot, and Google AI Overviews when live results aren’t available, are estimated with the model behind them and labeled that way.
The unit: Each prompt is asked once per engine, per region, per language, per scan.
The cadence: Daily, weekly or monthly on Pro and above. Weekly or monthly on Starter. Plus scans on demand.
The sharing: The web search behind a prompt runs once per prompt, region and language, and is shared across the engines in that scan. Competitors are read from the same answer your brand was read from.

How many runs stand
behind a number

A visibility score is a sample, and a sample has a size. Ours is one run per prompt, per engine, per region, per scan. Your score is the aggregate of those answers, tracked over time rather than read on any single day.

Why not twenty runs per question?

In research on answer variance, a repeat past the fifth barely reduces the error, while breadth across languages and models reduces it far more. We spend the budget on breadth, and on frequency over time.

Żatuchin, arXiv:2607.13304, Jul 2026. The study models sentiment polarity, not recommendations.

A weekly scan on Pro

200prompts×7engines×1run

1,400answers a week, for one region and language

Read as a trend across weeks, never as one week’s reading.

The same question doesn’t
get the same answer twice

This is the central problem of the category, and most tools don’t mention it.

Visibility across scansNoise bandYour score

noisenoisesignal

RoadmapA visible noise band in the product. This chart shows the principle. The band itself isn’t in the product yet, and we’d rather show you the thinking than wait.

What the research says

Share of the variance in one answer’s reading of a brand

Re-asking the same prompt34.8%

The language of the query26.5%

The brand itself1.5%

Żatuchin, arXiv:2607.13304, Jul 2026 (sentiment polarity)

“Single-run visibility metrics provide a misleadingly precise picture of domain performance in generative search.”Sielinski, arXiv:2603.08924 v3, Aug 2026

“Answers can vary across runs, prompts, and time, making one-off observations unreliable.”Schulte, Bleeker and Kaufmann, arXiv:2604.07585, Apr 2026

What we do about it

We never show you one answer and call it a measurement.
The score is a trend across scans, not a reading.
A move of a couple of points is noise. Read the direction over several scans, not the change since Tuesday.

What each word means here

There’s no standard in this category. Four tools can use “share of voice” for four different things. Here is ours.

Mentioned: Your brand name appears in the answer text.
Recommended: The answer puts you forward as a choice, not merely as a reference.
Position: Where you appear in the order of brands named. Not a rank out of ten: there is no page two.
Visibility score: How often you’re named across the prompts you track, over the window shown.
Share of voice: Your share of brand mentions across your tracked prompts, measured against the competitors you chose.
Citation: A source the engine linked or named in that answer. We don’t filter out citations to pages that don’t exist.
Sentiment: How the answer characterizes you. The weakest metric here; see the caveat below.

Share of voice is measured inside a pool you picked. Track three competitors and it’s a three-competitor measurement, not the whole market. Adding the brands AI actually names beside you makes it more honest, and the product surfaces those.

Sentiment is the weakest metric in this category. One study puts only about 1.5% of the variance in an answer’s sentiment down to the brand itself. Treat it as a flag to go and read the answer, not as a KPI.

What we don’t measure

In plain words, so nobody has to find out later.

We don’t see anyone’s conversations.We run the prompts you track, on a schedule. We have no access to real people’s chats, and no tool that claims otherwise should be believed.
We don’t see everything an engine has said about you.Only the answers to the prompts you track.
We can’t attribute every visit.Many AI-assistant visits arrive without a referrer and land in Direct, and published estimates of how many vary widely. Clicks from Google’s AI Overviews arrive as Organic Search, a separate problem often confused with this one. How AI Traffic handles it
We can’t tell you what a fix will do before you make it.We measure the trend afterwards and say what moved, including when nothing did.
We don’t model prompt volume.Where we show demand, it comes from your connected Search Console. Your own data, not an estimate.

What actually works, and
how confident anyone should be

We’d rather tell you the state of the evidence than sell you a certainty that doesn’t exist.

What the research says

The one factor with clean causal testing is whether the engines can fetch your page. In controlled probes, a 403 or an answer that only appears with JavaScript stopped retrieval outright. Adding llms.txt, JSON-LD, sitemaps or Markdown mirrors didn’t change whether an answer already in the HTML could be retrieved.

Vercel and ora.ai, Aug 2026 · 1,033 agent runs across 25 sites, 190 controlled fetch probes

Google says there are “no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.”

Google Search Central, AI features and your website

What we therefore do

1Make sure the engines can reach and read you. This is the part with evidence behind it.
2Make sure the page answers the question being asked, in the words buyers use.
3Measure the trend afterwards and report what moved, including when nothing did.

And what we won’t do

Sell you a small lift as a result. With this much variance between runs, a few points in one scan is weather, not a win.