How we measure, in plain words.

We ask the questions your buyers ask: “best tool for…”, “X vs Y”, “alternatives to…”. Questions are grouped by topic, so you can see where you show up and where you don't.

Each question carries two labels. A volume band (high, medium, low or unknown) is a rough guess at how often people search for it. A source says where it came from: your site, search data, your competitors, our templates, or you.

Six engines, in every plan. We ask each one the same questions, the same way, every few days, and keep exactly what it said.

On paid plans, every question is asked on every engine once every 3 days. Over a month that adds up to hundreds of answers per brand, which is what makes the ranges below narrow enough to be useful.

The free audit is a single snapshot: 5 questions on 4 engines (ChatGPT, Gemini, AI Overviews, Perplexity), asked once. That's 20 answers, enough to spot the obvious gaps, not enough to track change.

Every raw answer, kept as a receipt. Each number in ReachGraph links back to the answers it was counted from, so you can read them yourself.

First, plain matching. We look for your name, any aliases you give us, and your domain, and do the same for each competitor. That part is exact and repeatable.

Then a model reads each mention and judges two things: were you recommended or merely mentioned, and was the tone positive, neutral or negative. We store which version of this step produced each mention, so when we improve it you can tell old judgments from new ones.

Visibility

Answers that mention you, divided by all answers. We show it per engine and blended. The blend weighs each engine equally, so one heavily sampled engine can't drown out the rest. Next to every percentage you'll also see the plain count, like “mentioned in 7 of 20”.

Share of voice

Of all the brand mentions across your answers, the share that are yours. It shows who takes up the room when you're not there.

Citation share

Of all the links the engines cite, the share that point to your own site.

Sentiment

A simple label: mostly positive, mostly negative, neutral or mixed. We need at least 5 mentions before giving one; below that it says “not enough data”.

Position

Where in the answer you first appear: the first, middle or last third. That's all.

Every percentage comes with a 95% range: where the true number most likely sits, given how many answers we have. We use a standard method called the Wilson interval. Fewer answers means a wider range.

Both are 35%. With ten times the answers, the range shrinks from 18–57% to 29–42%.

We only say a number changed when this week's range and last week's range don't overlap. If they overlap, the honest answer is “not yet”, and that's what we say. Blended ranges are the average of the engine ranges, which is slightly cautious on purpose.

Anything based on fewer than 20 answers is marked low confidence.

Fixes come from six rules. Each looks for a specific pattern in the answers and links to the answers that triggered it.

Every fix is graded strong, moderate or weak. Each rule starts at the level shown above. If the finding rests on fewer than 20 answers, it drops one level. Two rules work a little differently: robots.txt is a direct check of your file, so it stays strong, and a wrong fact counts as strong once we've seen it in two or more answers.

Fixes are ordered by how many answers they could affect, how strong the evidence is, and how much work they take.

llms.txt. It's a proposed file that summarises your site for language models. No major AI engine has said it uses llms.txt when writing answers, so we don't sell it as a fix. If that changes, we'll add it and say why.

Answers vary by location, by account, and by time. We record where and when each answer was asked, but what you see when you ask ChatGPT yourself, signed in, somewhere else, on another day, may differ. That's expected.

We measure what engines say, not why. The fixes are our best reading of the evidence, graded honestly, not a promise.