How we measure

Last updated 8 October 2026

A Wellread audit asks ChatGPT and Gemini the questions your buyers ask, and reports who they name, how they present each company, and whether your website answers those questions. This page says exactly how each number is made, and where its limits are.

The questions

Wellread reads your website, works out who your buyers are, and writes the questions they would ask an AI assistant. You can add your own.
An audit asks up to 30 questions. Your own questions come first, within that number.
No question Wellread writes contains your company’s name. Ask an assistant about a company by name and it will talk about that company; that says nothing about whether a buyer who has never heard of you would be pointed to you. If you add a question of your own that names your company, it is asked and counted as you wrote it, and Settings warns that it will flatter your result.
One more question does name you: “What is [your company], and what does it offer?” Its answer is shown as how AI describes you. It is asked once to each assistant and is never part of any count.
Before anything is asked, you check what Wellread understood: your business and the questions.
Later audits of the same site ask the same questions, so two audits can be compared.

Who we ask

ChatGPT and Gemini, through their makers’ own APIs, with web search switched on. ChatGPT is told to search the web for every answer. Gemini decides for itself, and answers some questions from memory. Each report names the exact models it asked.
Search being on doesn’t prove an answer used it. So the report records, for every answer, whether it searched and which sites it drew on, and counts “used your site as a source” only over the answers that searched.
Each question is asked once to each assistant.
This is not the same as typing into the ChatGPT or Gemini app. Answers there depend on the person’s history, settings, location and the model they chose, so what you see in the app can differ.

What “named” and “recommended” mean

Every answer is read a second time by an AI model, against what your business sells, to decide whether it names your company. A different company that shares your name doesn’t count.
Named means the answer mentions your company. For each answer that names you, the report also records how: recommended first, given as one of the options, compared with others without a verdict, or advised against. Only the first two are a recommendation. The headline number counts answers that named you, and says beside it how many of those recommended you, so a neutral or negative mention is never shown as a recommendation.
The headline counts answers to questions where buyers look for or compare options, because those are the answers that name companies. On how-to and should-I questions, where assistants seldom name anyone, we count whether your site was used as a source.
An answer that names no company, product or method at all isn’t counted. Being left out of an answer that names nobody says nothing about you. The “Named by AI” figure explains how many answers were set aside this way, and the spreadsheet of every answer marks each one.
An ask that got no answer, or an answer that couldn’t be read against your business, is not counted as leaving you out. The report says how many there were.
Each answer is read, and kept in the report, up to its first 3,000 characters. A longer answer is marked as cut. A rival it names only after that point isn’t counted; your own company still is, if the answer names it later and cites your site.
An assistant sometimes answers from memory without searching the web. Such an answer can name companies, so it still counts toward who is named. It cannot cite any site, so it isn’t counted where we measure whether your site was used as a source.
Your rivals are the companies the answers actually named. Those named in 3 or more answers are marked when an audit finishes, and you choose the list from there. They are never guessed from your name.
For the rivals named most, the report quotes a sentence about each one, from the answer that recommended it most strongly. These are the assistant’s words; Wellread doesn’t write a reason for it, and shows no quote where the answer only listed the name.

What “your site answers” means

Wellread reads up to 60 of your pages the way an AI crawler does: without running JavaScript. If that gives almost nothing, it reads up to 15 pages as a browser shows them, which is what Google can use.
For each buyer question it says whether a page answers it, partly answers it, or doesn’t. “Answers it” needs a quote that is actually on the page.

When a number changes

AI answers vary from one run to the next. A change of 2 answers or fewer is shown as “about the same”, not as up or down.
Two audits are compared only when they checked the same site, asked the same questions, were counted by the same rules, and the same assistants answered.
If too few of one assistant’s answers come back (under 80 in 100), that assistant is left out of the audit and the report says so: a result built on a few of its answers would mislead. The other assistant’s answers still make a report. If neither comes back, the audit stops and asks you to try again.

When we can’t read a site

Some sites answer automated visitors with a bot check, or block visits from outside their country. Our reader runs in India.
If Wellread could read almost nothing of a site (under about 200 words, counting repeated text once, so the same error page at twenty addresses counts as one), the audit stops and says so. Nothing is scored, and it doesn’t use one of your audits.
The “can AI crawlers get in” checks are asked from our server using each crawler’s name. The real crawlers come from their own addresses, so confirm in your CDN or firewall settings before changing anything.

What we don’t measure

What Claude, Perplexity or Copilot say. We check whether their crawlers can read your site, not their answers.
Personalised answers inside the apps.
Traffic, rankings or sales that come from AI.
A statistically stable score. An audit is a snapshot of 30 questions on one day; its use is to show the pattern, who gets named and what your site is missing, and to check again after you change something.