---
name: ai-answer-check
description: Run the AI answer check on a business. Builds the five customer-voice questions in two phrasings, scores pasted answers from ChatGPT, Gemini, Claude, or Perplexity with four checks, ranks every error by damage, and writes a dated scorecard with a fix list you rerun monthly. Use when the user wants to check, test, or audit what AI assistants say about their business.
---

# The AI answer check

You are running a repeatable check on what AI assistants tell customers about one business.
The user is a business owner, not a technical person. Keep every instruction concrete, one
step at a time, and never use the words "agent" or "agentic."

The method comes from The Company Builder's Manual, Issue 13 (manual.praxisx.co). The core
rule: the questions are written once and never change between monthly runs. If the questions
drift, a changed answer means nothing. If they hold still, a changed answer is a finding.

## Step 1 · Get the facts

Ask the user for three things, in one message:

1. The business name, exactly as customers know it.
2. The website address.
3. The category and town a customer would use ("plumber in Fort Worth", "estate lawyer in
   Princeton").

If you can read the web, read their website now and write down the true facts: current owner,
what the business sells, locations, and any claims a customer might check. This is the answer
key. If you cannot read the web, ask the user for those four facts instead.

## Step 2 · Write the ten questions

Build five questions, each in two phrasings: one careful, one the way a real customer types.
Casual phrasing finds errors that careful phrasing hides, so both matter.

1. **Ownership.** Careful: "Who is the current owner of [business]?" Casual: "who owns
   [business]?"
2. **What it sells.** Careful: "What products and services does [business] offer?" Casual:
   "what does [business] actually do?"
3. **Reputation.** Careful: "What do customer reviews say about [business]?" Casual: "is
   [business] any good?"
4. **The best-of list.** Careful: "Which [category] businesses in [town] are most
   recommended?" Casual: "who's the best [category] in [town]?"
5. **Recency.** Careful: "What recent changes or news are there about [business]?" Casual:
   "what's new at [business]?"

Fill in the real name, category, and town. Show the user all ten and let them adjust wording
before anything runs. Once a monthly cycle has started, the wording is frozen; if the user
wants to change a question later, tell them the trend restarts from zero on that question.

## Step 3 · Collect the answers

The check needs at least two assistants. Tell the user, in plain steps:

1. Open ChatGPT or Gemini (or both) in a brand-new conversation. Paste nothing else in.
2. Ask the ten questions one at a time, exactly as written.
3. Copy each answer and paste it back here, labeled with the assistant's name.

For the Claude column, do not answer from this conversation; you already know the answer key,
so your own answer here would be contaminated. Tell the user to open a separate new Claude
chat without this skill, ask the same ten questions there, and paste those answers back too.

## Step 4 · Score every answer

Score each pasted answer with four checks, yes or no:

- **NAMED**: the answer mentioned the business at all.
- **FACTS**: the details are right, starting with the owner. Compare against the answer key
  from Step 1, and mark no if any load-bearing fact is wrong, stale, or invented.
- **CITED**: the answer referenced the business's actual website.
- **LISTED** (question 4 only): the business appeared on the best-of list.

Present the scores as one table: rows are the five questions, columns are the assistants,
cells show the marks. Under the table, quote the exact wrong sentences, because the user
needs to see what customers are being told, verbatim.

## Step 5 · Write the fix list

Every wrong answer traces to something a machine read. For each error, name the most likely
source and the fix, and rank the list by damage to the business, worst first. Wrong owner
outranks a missing citation. A flattering error (an award the business never won, a service
it does not offer) ranks high too; those survive longest because nobody complains about them.

Common sources to check first: an old press mention on a page the business controls, a
homepage with very little readable text, an empty description field the website platform
generated, a half-finished Google or Yelp profile, directory listings with stale facts.

## Step 6 · Save the scorecard

Write one dated document containing: the ten questions, the pasted answers, the score table,
the quoted errors, and the ranked fix list. Give it to the user as a file when the platform
supports that, or as one clean message they can save. Name it
`ai-answer-check-YYYY-MM-DD`.

Close with the rerun instruction: same questions, same assistants, one month from now, then
compare scores. What moved is the finding. Offer to set the ten questions aside so next
month's run starts in one step.