Answermend Sign in Check my site

How it works

How we check whether AI answers name your business.

When one of your customers asks ChatGPT or Gemini something like “which brands make the best waterproof hiking boots”, the reply names a handful of brands. Either you are one of them, or you are not. This page walks through exactly how we find that out.

Open the free tools What the four results mean no card
Your business yourbusiness.co.uk EXAMPLE DATA

12 of 18

Named in

measured answers

5

Cited

page opened + linked

7

Mentioned only

from model memory

2

Not measured

excluded from counts

Example data. Five questions, asked once each in ChatGPT and in Gemini; four rows shown. Two checks did not return, and are left out of every count.
Question ChatGPT Gemini Named instead
Best review software for contractors cited cited —
Top review management providers mentioned missing competitor-one.com
How much does review software cost? missing mentioned competitor-two.com
Is yourbusiness.co.uk any good? not measured cited —
Checks the answers in ChatGPT Gemini Claude Pro plan Perplexity Pro plan Google AI Overviews has no public API and is not measured — assistants we cannot measure are named, not guessed

What we actually do

Three things. That is the whole idea.

Whole sentences, asked the way your customers ask them — then one honest line written down per question.

  1. 01

    We write down the questions your customers would ask

    Whole sentences, the way people really type them — “best family solicitor in Bristol”, “is [your business] any good” — not single words.

  2. 02

    We ask those questions and read the answers

    We put each question to the AI assistants your customers use, and read what comes back.

  3. 03

    We write down one thing per question

    Were you named — and if you were, did the assistant actually open your website, or was it going from memory, the way a person half-remembers a shop?

The question sheet yourbusiness.co.uk EXAMPLE DATA

01best crm for small law firms

02top review management providers

03how much does review software cost?

04is yourbusiness.co.uk any good?

five of these per scan · written as sentences, never as keywords

Inside one answer

It either opened your website, or it did not.

An assistant naming you from memory and an assistant opening your website and quoting it are two different events with two different causes. Only one of them is something your website can be credited or blamed for — so we never add the two together.

One question, traced yourbusiness.co.uk EXAMPLE DATA

Your customer asked

“best crm for small law firms”

ChatGPT

What it did

Opened yourbusiness.co.uk while writing the answer

A page of yours was fetched, and the answer pointed back at it.

So the result is

cited1 of 1 run

Your own page did the work, and we can point at the page that did it.

Gemini

What it did

Answered from memory — nothing of yours was opened

Your name came out of the model, not out of your website. We cannot see inside that.

So the result is

mentioned1 of 1 run

Worth knowing — but no page of yours caused it, so no page of yours can be credited for it.

Then the mend

The gap becomes a file you can paste, or forward to whoever built your site.

faq-schema.json an answer section on the closest page llms.txt entry

We ship the file and ask the same question again. We do not promise a mention — nobody controls what a model answers.

one question · two assistants · one run each

Solid — your website was opened, so there is something real to point at, and something real to fix.

Dashed — it was not opened. That is a gap we cannot see into, and we draw it as a gap rather than filling it in.

Step by step

What happens when you run a check

Four steps, in the order they happen. You give an address, the questions get asked, your own pages get opened, and what comes back is a list plus the files that fix it.

  1. 01

    You type in your website address

    yourbusiness.co.uk

    the address people type to find you — that is all we need to start

    No plugin, nothing added to your site — nothing changes unless you change it yourself.

  2. 02

    We ask the questions your customers ask

    “best crm for small law firms?”

    Free check: up to 5 questions, in ChatGPT and in Gemini, one pass each. Paid plans: up to 10 questions on Starter and 40 on Pro. Both ask each question once per ask and come back the next day, so the repeats accumulate across the month rather than arriving together. They are reported as a spread carrying their own run count, with the history kept so you can see what moved. Starter asks the same two assistants; Pro adds two more — Claude and Perplexity. Both run daily — that scheduler has run every day since 6 Sep 2026, so far for our own Pro account only. A re-check is also a button you press, on any plan.

    Every answer is recorded as one of four states, and a question we could not get an answer to is recorded as not measured rather than as a miss.

  3. 03

    We look at your own website the way an assistant does

    Are AI companies allowed in?Sites can accidentally tell them to stay out. It is one line, in one small file.

    Can they read what is on the page?A page that has to load extra code before its words appear can look blank to them.

    This half never asks an assistant anything — it only opens your own pages. All three checks that do this — Crawler Checker, Readiness Score and Render Check — run today, on our server rather than in your browser.

  4. 04

    You get the results, and the files that fix what we found

    faq-schema.json open ↗ robots.txt open ↗

    Results come back as a plain list — question by question, what happened. Fixes come back as files. The four that make files are free and open right now.

app.answermend.com EXAMPLE DATA
Question-by-question table for Northlight, one column each for ChatGPT, Gemini, Claude and Perplexity: cells read mentioned, cited, missing or not measured, each saying what happened on its one run, and a final column lists the domains named instead.
The results screen as it renders. No cell prints a bare verdict: each chip says what happened on its one run, two questions — “Top bookkeeping companies” and “Most affordable bookkeeping companies?” — are missing on all four assistants, and a Perplexity check that was rate-limited is printed as not measured, never folded into a count. Northlight is an invented brand; its scan is simulated.

We would rather show you an empty box than a made-up number. Where a check could not reach something, the page says so.

The four results

What the four results mean

Every question in your report lands on one of these four. There is no fifth.

  • cited

    Your page was used as a source

    The assistant opened your website while writing the answer, and pointed at it. This is the strongest of the four: your own page did the work, and there is a link back to you.

  • mentioned

    You were named from memory

    Your name came up, but nothing of yours was opened and nothing was linked. Worth knowing — but no page of yours caused it, so no page of yours can be credited for it either.

  • missing

    You were not in the answer

    Your business did not appear at all. Somebody else has the slot, and the report tells you who. This is the one worth acting on first.

  • not measured

    We could not check this one

    The question never actually got asked — a check failed, or there is no honest way for us to reach that assistant. Nothing came back, so nothing is recorded against you.

app.answermend.com EXAMPLE DATA
Four report tiles in one row: 6 of 8 prompts named you, 4 of 8 cited you, 6 of 8 mentioned you, and a dashed tile reading 2 checks not measured of 32 prompt-by-engine checks, excluded from all three denominators and never counted as a miss.
The four states on the report itself: named on 6 of 8 prompts, cited on 4 of 8, mentioned on 6 of 8. Each tile prints its own denominator in small type — “prompts where an engine answered” is a definition, not decoration — and the dashed fourth tile counts the checks that did not run, excluded from all three denominators rather than scored as misses. Northlight is an invented brand; its scan is simulated.

Why “we could not check” is never shown as a zero

A zero means we asked and you were not there. Not measured means we never asked. Those are two different facts, and the report never prints one as the other: a number on your report was earned, or it is not there at all.

EXAMPLE DATA

What the report prints

not measuredwe did not ask

What it never prints

0 of 5 — struck through: never printeda count nobody earned

Repeated runs

Why we ask the same question more than once

Ask an AI assistant the same question five times and you can get five slightly different answers. That is not a fault, it is simply how they work.

Which makes a single check close to a coin toss. So we ask each question several times and report the spread — named in three runs out of five — rather than one lucky answer written up as a fact.

On the free check the rule is one pass per question, labelled as one run rather than as an average. Asking repeatedly, and watching the number move month to month, is what the paid plans are for.

One question, asked five times EXAMPLE DATA

Runs

1named

2named

3named

4not named

5not named

3 of 5 runs named this business. Two did not. The honest way to write that down is a spread, not a yes and not a no.

Solid — the run came back, and it named you.

Dashed — the run came back, and it did not.

Illustrative example. Not a real scan.

The rest of the record

Four more things the check writes down

Each is its own screen, and each says how much of it was really measured.

  • Who got cited instead of you

    When an assistant answers and never opens your site, something else was used to build that answer. The report lists those pages — the reviews, directories and comparison sites it drew on — so you can read what they say that you do not. It is built only from answers that were really measured: a run nobody made is left out, and the leaving out is counted rather than quietly dropped. Starter and up.

  • What the assistant searched for first

    Before answering, an assistant often runs its own searches. Where the engine tells us what it typed, those words are on the report, stated against how many checks reported any — never as a bare list. Where it searched and did not say what for, that is what we record: not a search that found nothing, and never a zero. On the free plan only Gemini opens your site, so on ChatGPT there is nothing to capture.

  • Google’s own AI answer

    Whether Google shows an AI Overview for your questions, and whether it names you or uses your page. It is a results page rather than an assistant’s reply, so it is kept apart: never added into a scan’s counts and never averaged with them. Nothing is asked until you choose the questions, because each one costs a search at a paid vendor. Pro.

  • When an answer moves

    The same question, asked again later, can come back different. An alert tells you when it does and what changed — and a difference has to survive the spread between repeated runs of the same check before it is called movement rather than noise. Pro.

The mend

What you get afterwards

Files, not homework.

  • faq-schema.json open ↗

    Your common questions, in a form assistants can quote

    You paste one block of text into your site. It sets out your questions and your answers in the tidy, labelled form these systems read best.

  • organization-schema.json open ↗

    Who you are, stated plainly

    Another block of text to paste in: your real name, where you are, your official profiles. It stops assistants mixing you up with a similarly named business somewhere else.

  • robots.txt open ↗

    Who is allowed to read your site

    A small file you upload to your site, listing which AI companies may read it and which may not. Written correctly, so you do not lock out the ones you wanted in.

  • llms.txt open ↗

    A short summary of your site

    One more small file to upload — a plain summary of what your site holds. Worth the hour it takes, though no AI company has confirmed it reads this file, and we are not going to pretend otherwise.

Those four you can make right now, free. They run in your browser from what you type in, so nothing about your business is sent to us.

What you will not get is a list of jobs to go and do. If a thing can be written for you, we write it and hand it over. Paste it in yourself if you are comfortable doing that — or forward the file to whoever built your site, where it is a few minutes of work.

faq-schema.json paste into your page EXAMPLE DATA

{ "@type": "FAQPage",

  "mainEntity": [{ "@type": "Question",

    "name": "How much does review software cost?" }] }

written from what you typed · nothing about your business leaves the tab
app.answermend.com EXAMPLE DATA
One What-to-fix item, chipped gap: questions answered without this brand in the answer. Do this: check that each question has a page answering it in its own words, where an AI crawler can reach it. Because: the brand was neither named nor cited in 2 of the 8 answered questions. Then the two questions, and a button to record a fix.
One item from the report’s fix list. It says what to do, why (neither named nor cited in 2 of the 8 answered questions), and names the two questions it rests on, so the instruction traces back to a measured answer rather than to a generic checklist. Northlight is an invented brand; its scan is simulated.

What we do not do

Three things we will not sell you

If anyone offers you any of these, ask how they intend to keep it.

  • No guaranteed mentions

    Nobody controls what an AI assistant says about anybody — not us, not the assistant’s owner, and certainly not a monthly subscription. Anyone promising you a mention is selling something they cannot deliver.

    What we promise

    • The checks run, and you see every result, including the ones that went badly
    • The files arrive, ready to paste or upload

    What we never promise

    • That you will be named in any answer
    • That your page will be used as a source in any answer

  • No ranking claims

    An AI answer is not a numbered list you can climb. It is a paragraph, written fresh each time, that either names you or does not. So there is no position to buy, and we do not describe our work as moving you up one.

    What a result says

    citedyour page was opened and used

    What it never says

    position 3 — struck through: never printeda rank that does not exist

  • No score out of 100

    A single tidy number feels good and tells you nothing. It hides which questions you lost, which assistant lost them, and how many checks ran at all. We report what happened instead, even when that is less comfortable to read.

    What the report prints

    named in 3 of 5 runswith the two that failed shown

    What it never prints

    82 / 100 — struck through: never printedone number hiding all of it

app.answermend.com EXAMPLE DATA
Sentiment across four assistants for the first two questions: counts of runs, never a score, and a cell is published only from 2 measured runs or more, so each cell here reads what its one run said — positive or neutral, not published; Claude did not name the brand on the first question; Perplexity's second cell is not measured.
One run is not a distribution, so the screen publishes a cell only from two measured runs or more; below that, each cell says what its one run said. The Perplexity check that did not come back is marked not measured — never neutral, never a zero. Northlight is an invented brand; its scan is simulated.
app.answermend.com EXAMPLE DATA
The side card on the sentiment screen: this scan's run and a link across recent scans; no paid call was made for this view, because the labels were written when the scan ran, by openai/gpt-4.1.
Every report view says what it cost to draw. The classifier is named, and this view made no paid call — the scan wrote the labels when it ran. Northlight is an invented brand; its scan is simulated.

Fair questions

Short answers.

One row, in full EXAMPLE DATA

Top review management providers

ChatGPTmentioned1 of 1 run

Geminimissing1 of 1 run

Insteadcompetitor-one.com

every cell says how many runs returned, and how many held

Do I need to be technical to use this?

No. The results are written in plain sentences, and the fixes are files you can forward to whoever built your site. Nothing is installed, and nothing on your website changes unless you choose to paste something in yourself.

What is the difference between being mentioned and being cited?

Mentioned means an assistant named you from memory, with nothing of yours opened and nothing linked. Cited means it actually opened your page and pointed at it as a source. Different causes, different fixes — so we never add them together.

Why does it say “not measured” instead of zero?

Because those are two different facts. Zero means we asked and you were not in the answer. Not measured means the question never got asked, so there is nothing to count and we will not invent one.

What can I use without paying?

The four tools that make files for you are open and free right now. So are the three that look at your own website — the crawler check, the readiness score and the render check: you give one a page address, our server fetches it, and you get a real result back.

Can you guarantee my business gets mentioned?

No, and nobody can. What we can do is show you where you are missing, tell you who is there instead, and hand you the files that fix what those systems actually read on your site.

Start with the parts that run today.

Four file makers and three site checks, open on the shelf.

Open the free tools → no card
faq-schema.json organization-schema.json robots.txt llms.txt