How to read your numbers

Every figure in WhatGPTSays comes from the same loop: we ask AI assistants a fixed set of realistic questions on a schedule, record each answer, and count. Below is what each number means, how it is computed, and what it does not tell you.

  1. Measure: look at Overview. Is your share of voice up or down against the previous period, and on which engine?
  2. Diagnose: open Prompts. Which questions do you miss, and why (no content, not cited, missing from sources, negative tone)?
  3. Place: open Sources and Categories. Which sites do AI engines cite for your market, and which of them can you get listed on?

AI visibility

How often and how favourably AI assistants mention your brand and point to your pages when people ask questions in your market.

AI assistants (ChatGPT, Perplexity, Google AI, Gemini, Claude) answer questions directly instead of listing links. Your AI visibility is whether your brand appears in those answers, how it is described, and whether your own pages or pages about you are cited as sources. It is measured by asking the assistants a fixed set of realistic questions (prompts) on a schedule and recording what comes back.

Prompt

A question a real person might ask an AI assistant. We ask the same prompts every week so changes are comparable.

A prompt is one tracked question, for example 'best CRM for B2B startups'. Every prompt has a type: category (searching by need, without knowing brands), comparative (comparing named products), problem (a how-to or job to be done) and brand (asking directly about a company). The type is kept everywhere so you can see which kind of question you win or lose.

Prompt type

Category, comparative, problem or brand: what kind of question it is.

Category prompts test whether you appear when people do not know brands yet. Comparative prompts test whether you are included in comparisons. Problem prompts test whether you surface as the solution to a task. Brand prompts show what AI says about you when asked directly. Most growth comes from category and problem prompts.

Answer

One engine's reply to one prompt in one run. All percentages are computed over answers.

If you track 20 prompts on 4 engines and run once, you get 80 answers. An answer can mention you or not, can cite sources or not, and has a tone. Rates are 'answers that ...' divided by 'answers', which is why the number of answers is always shown next to a percentage.

Share of voice

The share of AI answers that mention your brand. 30% means 3 of every 10 answers named you.

Share of voice = answers that mention your brand divided by all answers in the period, from 0% to 100%. It answers 'how visible are we?'. A change of a few points between two weeks can be noise because AI answers vary from run to run; look at the trend line and at the previous period, not one point. It says nothing about whether the mention was positive: see sentiment.

Mention

Your brand name or domain appears in the text of an AI answer.

A mention is counted when your brand name or the name part of your domain appears as a whole word in the answer text. Being mentioned is different from being cited: an answer can name you without linking to you, and can link to you without naming you.

Citation

An AI answer lists a page as one of its sources. A citation of you means one of your pages is a source.

Many assistants attach the web pages they used. Each such page is a citation. 'Citation rate' is the share of answers that came with sources and cited your own domain. Not every engine returns sources (Claude, and sometimes Gemini, do not), and the app says so instead of showing an empty list as if nothing was cited.

Source

A web page or site an AI answer is built from. Sources are what you can act on.

If the engines keep citing the same review sites, media and forums for your market, getting listed or mentioned there is the most direct way to appear in answers. The Sources screen ranks domains and individual pages by how often they are cited, with what kind of site they are, and which of them you can place content on.

Source type

What kind of site a source is: media, brand, user-generated, review or reference.

Types are assigned by simple rules, not guessed: UGC / social (Reddit, YouTube, LinkedIn ...), reference (Wikipedia ...), review sites (G2, Capterra ...), catalog media (sites in our media catalog, where a placement can be bought) and other (company sites and everything we do not classify). 'Other' is deliberately not split further.

Sentiment

The tone of the answer around your brand, from negative (0) to positive (1). Rule based, not a model's opinion.

Sentiment is scored with a fixed list of positive and negative phrases in the text near your brand name, so the same text always gets the same score. It is a coarse signal: use it to spot a change, then read the answers behind it.

List position

Your rank inside numbered or bulleted lists in AI answers (1 = first).

When an answer is a list of options, being first matters. List position is the average rank of your brand across the answers where it appears inside a list. It is empty when you never appear in a list.

Engine

The AI assistant that produced the answer: ChatGPT, Perplexity, Google AI, Gemini or Claude.

Engines differ: some search the web and show sources (ChatGPT, Perplexity, Google AI), some only partly (Gemini) and Claude answers from training data. The per-engine view shows where you are strong and where you are absent. An engine that failed to answer during a run is shown as missing data, never as zero.

Change vs previous period

The difference between this period and the equally long period just before it.

Overview numbers cover the last N days (default 30). The change compares them with the N days before that. A green or red arrow shows direction only when both periods have answers; otherwise a dash is shown. Differences smaller than about 5 points on a few dozen answers are within normal variation.

Category index

A fixed set of about 100 questions for one market, asked to the engines every week, to show who AI cites in that market.

The index is our own panel, independent of any customer: for each category (fintech, health, SaaS ...) we ask the same questions weekly, including 'where to publish a press release in this market'. It shows the most cited domains and pages, which source types dominate, and who entered or left the top. It contains no customer data.

Where to get listed

Sites that AI engines cite for your category and where a placement can be bought, ranked by how often they are cited.

This list joins two facts: the engines cite the site for questions in your category, and the site is in the media catalog. It is a ranking by citations, not a promise: a placement makes it more likely, not certain, that an assistant cites you.

Diagnosis

Why you are missing in an answer: no content, not cited, missing from sources, or negative tone.

Diagnoses come from fixed rules comparing the answer with your content and with the sources it cited. The same input always gives the same diagnosis. Each diagnosis maps to concrete tasks in the action queue.

Before / after comparison

Compares the prompts a placement targeted with your other prompts, before and after, to separate the effect from general movement.

We need at least 3 runs and 30 answers in each window. The result is the change on the targeted prompts minus the change on the control prompts, with an interval. It is labelled directional: AI answers vary and one placement is one cause among many.

Questions about a number? Ask us. Developers: API reference.