Guide 3

Does ChatGPT recommend your business? Here's how to test it yourself

The guide from the video, with the full AI instruction ready to copy.

1. What you'll end up with

You ask ChatGPT the questions your customers ask when they're looking for someone like you. Each one three times, in a fresh chat every time. A second AI acts as your test lead: it writes the questions with you and then counts what's in the nine answers. At the end you have:

three test questions in your customers' words, ready to reuse,
a count from nine ChatGPT answers: how many name you, who gets named instead of you, which websites ChatGPT shows as sources,
a rating for each question: mostly named, sometimes named, not named,
five things at most to look at next, each tied to the place in the answers it's based on.

Before you start: the result is a snapshot. AI answers aren't fixed. The same question can bring different names tomorrow, and another customer somewhere else may get a different answer. That's why you ask each question more than once, and why you repeat the test later with the same questions.

In the video you saw our system putting these questions to ChatGPT and Perplexity on a regular schedule. This guide is the by-hand version: the same idea, with the ChatGPT app and a second AI chat.

Who it's for: business owners, the self-employed, agencies, consultants and service providers whose customers ask, "Who should I hire?" No coding needed.
Time: set aside a quiet 45 minutes the first time: answer a few questions, run nine short test chats, read the results.
What you need:
- ChatGPT (chatgpt.com or the app). The free version is enough: according to OpenAI's help pages, ChatGPT can search the web on the free plan, even without signing in.
- a second AI chat as your test lead, for example Claude (claude.ai). Claude has a free plan; according to its pricing page, every plan has usage limits that reset in a five-hour window. You can also use ChatGPT as the test lead, but then in its own temporary chat (step 3).
- your business name in every spelling people use (with and without "Ltd" or "LLC", the short form, an old name) and your website address.
- optionally Perplexity (perplexity.ai), an AI search engine that shows its sources with every answer. According to Perplexity's help pages, basic searches on the free plan are "practically unlimited".

2. Step by step

Part A: set up the test lead

Why not just ask in your normal ChatGPT? ChatGPT can remember things from earlier chats (the feature is called "Memory"). According to OpenAI, saved memories can even shape what ChatGPT searches the web for. If you've told ChatGPT about your business before, you may not get the answer a stranger would get. So in part B you ask the test questions in a temporary, unpersonalized chat.
Open the test lead chat and make sure what you type isn't used for training.
- Claude: go to claude.ai/settings/data-privacy-controls and turn off "Model Improvement". Or use an Incognito chat: according to Anthropic, those aren't used for training.
- ChatGPT: your profile → Settings → "Data controls" → turn off "Improve the model for everyone".
If you also use ChatGPT as the test lead, open a temporary, unpersonalized chat for it (as in step 7). According to OpenAI, a temporary chat doesn't create memories, so nothing about your business from the test lead conversation can slip into the test chats later. Keep the window open: once you leave a temporary chat, it isn't in your history.
Copy the AI instruction from section 3 into your notes app and fill in the block "What you already know about me". Replace everything in [SQUARE BRACKETS] with your details. Leave blank what you don't know, and the AI will ask. You need the date because the result only holds for that day.
Paste the whole instruction as the first message in the test lead chat and answer the questions. Six at most, one at a time. Short answers are fine.
- "skip" skips a question.
- "done" ends the conversation right away. You get the results with whatever is there so far.
You get three test questions and a test plan. Copy the three questions into your notes. Check: is your business name or website in one of them? Then tell the test lead, because otherwise the test is worthless.

Part B: the nine test chats

In ChatGPT, open a temporary, unpersonalized chat. According to OpenAI's help pages: open a new chat → select "Temporary" → choose "Unpersonalized" before your first message. You can't change it after that. An unpersonalized chat doesn't use memory or custom instructions.
- No account works too: open a private browser window, go to chatgpt.com and don't sign in. According to OpenAI, you can turn off "Improve the model for everyone" under "Settings" there as well; it applies to that browser.
Paste question 1 word for word and send it. Add nothing, ask no follow-up. Don't switch on web search yourself, so you ask the way someone who just types a question would. According to OpenAI, ChatGPT searches the web on its own when the question would benefit from it.
Copy the whole answer. If ChatGPT shows a "Sources" button underneath, open it and copy the list of websites too. Copy everything before you leave the chat.
Paste the answer into the test lead chat with its label in front: Q1-R1: (question 1, run 1). The test lead confirms in one line and tells you which label is next. If it asks whether a similar name is your business, answer briefly.
New temporary chat, same question: Q1-R2, then Q1-R3. Then question 2 (Q2-R1 to Q2-R3) and question 3 (Q3-R1 to Q3-R3). A new chat every time, unpersonalized every time.
After the ninth answer you get the results. Out of time? Type "done", and the test lead counts what's there and says how many runs each question had.

Part C: read it and keep it

Spot-check one line: pick a row in the table "Every answer" and compare it with the answer you pasted. If it doesn't match, tell the test lead in the same chat and it recounts.
Save the result with its date in a document, together with the nine answers. You need both when you repeat the test (follow-up d).
Optional: the same test in Perplexity (follow-up a).

About location: according to OpenAI, ChatGPT may work out your rough location from your internet connection, and a VPN (a connection routed through somewhere else) can change it. That's why your test questions always name the city or region themselves. OpenAI suggests the same when local results show the wrong area.

3. The AI instruction, ready to copy

Copy everything in the box, fill in the placeholders, and paste it as the first message in the test lead chat. Not into the test chats: those only ever get the test question.

You're running a visibility test with me: we're checking whether ChatGPT names my business when customers ask for a provider like me. You don't ask the test questions yourself and you don't answer them. You help me write three good customer questions. I ask them in ChatGPT, paste the answers in here, and you count and report.

== WHAT YOU ALREADY KNOW ABOUT ME ==
(Ask about anything left empty. Don't ask again about anything filled in.)
- My business name and other spellings: [FOR EXAMPLE: MILLER DESIGN, ALSO "MILLER DESIGN STUDIO"]
- My website: [FOR EXAMPLE: MILLERDESIGN.COM]
- What I offer, in my customers' words: [FOR EXAMPLE: WEBSITES FOR MEDICAL PRACTICES, TAX HELP FOR THE SELF-EMPLOYED]
- Where my customers are: [CITY, REGION OR "NATIONWIDE, ONLINE"]
- Competitors I know of: [NAMES OR "NOT SURE"]
- Where I'm testing: [FOR EXAMPLE: CHATGPT, TEMPORARY CHAT, UNPERSONALIZED]
- Date of the test: [DD/MM/YYYY]

== HOW TO RUN THE CONVERSATION ==
1. Explain in two sentences what we're doing. Tell me: 6 questions at most, then three test questions, then I paste 9 answers. "not sure" is fine, "skip" skips a question, "done" takes us straight to the results. Then ask the first question.
2. Always ask exactly ONE question per message. With each question, give two or three short example answers from my line of business.
3. Order:
   a) Ask about whatever is missing in the block above.
   b) "How do new customers first ask about you, on the phone or by email? In their words."
   c) "What makes your customers choose one provider over another?" (for example price, experience with their industry, being close by)
4. Then suggest three test questions in one message and ask only: "Do these work for you?"
   - Q1 Recommendation: Who would you recommend for [service] in [place]?
   - Q2 Cost: Roughly what does [service] cost in [place], and who should I ask?
   - Q3 Choosing: What should I look for, and which providers would suit me?
   Each one sounds like a customer who doesn't know me, using their words from b) and c). No business name, no website, no competitor in the question. 30 words at most.
5. Then the test plan as a numbered list: ask each question three times, in a new chat every time, set up as in the block above. Paste the question word for word, add nothing, ask no follow-up. Copy the whole answer, including the list of sources if one is shown. Put the label in front of each answer: Q1-R1, Q1-R2, Q1-R3, Q2-R1 … Q3-R3.
6. When I paste answers, reply with one line per answer: "[label] logged: named were [names in order]. You: yes (position X) / no / source only. Sources: [number]." Then: "Next: [label]." No analysis yet. If the label is missing, ask for it. If a name looks like mine but you can't be sure it is, ask me and don't count it until I answer.
7. If something is unclear, ask once. Never guess. No compliments, no summary after every answer.

== HOW TO COUNT ==
- Named = my name or one of my spellings appears in the answer text as a provider.
- Source only = my website appears among the sources or as a link, but my name isn't in the text. That does NOT count as named.
- Position = where I come among the providers named in that answer (1 = first).
- Other providers = every other business or person named as a provider. Platforms and directories are not providers; they go under sources.
- Sources = websites from the source list or from links in the text, counted per website, at most once per answer.
- If an answer names no providers, it counts as a run with "no names".
Rating per question (runs = answers pasted for that question):
- "Mostly named" = named in more than half the runs
- "Sometimes named" = named in at least one run, but not in more than half
- "Not named" = named in no run
- Under 3 runs, add: "(only [N] runs)".

== FIXED RULES ==
- Count only what is literally in the pasted answers. Add nothing from your own knowledge: no business, website or review that isn't there.
- Don't invent numbers. Every number is a count from the pasted answers. No search volumes, no market figures, no percentages about customers.
- Don't claim how ChatGPT or any other software chooses providers or what it can do. Mark guesses "to be checked".
- A test is a snapshot. Answers change from run to run and over time. Never write that ChatGPT "always" answers this way or that the result is permanent.
- No promises like "then ChatGPT will recommend you", "guaranteed" or "fully automatic". Whether a step changes anything only shows in the next test.
- Don't judge competitors. Only report what the answers say about them.
- Don't answer legal questions (advertising, reviews, data protection, professional rules). Mark them "check with an expert". Never suggest bought or fake reviews.
- If I paste personal data about customers or staff, tell me once that it doesn't belong here.
- Write so that someone without any tech background understands every word. A technical term only with an explanation of five words or fewer.

== WHAT THE RESULTS LOOK LIKE ==
# Your AI visibility test
First line: "Snapshot from [date], tested in [where I'm testing], [number] answers counted. The same question can bring different names tomorrow."
## 1. Your three test questions
Word for word, so you can reuse them.
## 2. Every answer
Table: Label | You named? (yes/no/source only) | Position | Other providers in order | Sources
## 3. Result per question
Table: Question | Runs | You named (x of y) | Rating | Named most often
Underneath, one line: "Named overall: x of y answers. Source only: z."
## 4. Who gets named instead of you
Table: Provider | in how many answers | Labels | What the answers say about them (short, word for word where possible)
## 5. Which websites show up as sources
Table: Website | in how many answers | Type (your judgement: directory, review site, press, a provider's website, other) | Are you listed there? (always "to be checked", unless it's my website)
## 6. What to look at next
5 points at most. Each point: one observation with labels as evidence, then a step I can check myself. For example: a website keeps showing up as a source → check whether I'm listed there with current details; only others get named for one question → check whether my website answers that question; the answers praise something in others that I offer too → check whether my website says so clearly. If a step touches prices, reviews or advertising claims, add "check with an expert".
## 7. How to repeat the test
Same questions, same setup, for example in four weeks. Only compare tests with the same setup. With 3 runs per question, a difference can be chance.
## 8. What's still open
Unclear names, missing runs, your own judgements.

== BEFORE YOU SEND THE RESULTS ==
Check every point. If one isn't met, fix it first:
- Every count traces back to a label, and the totals add up.
- No provider and no website in the tables that isn't in a pasted answer.
- "Source only" isn't counted as named. Unclear names aren't counted.
- The rating follows the rules above exactly.
- The snapshot line is at the top. No sentence says the result is permanent.
- No sentence promises an effect or claims how ChatGPT chooses.
- Every point in section 6 has a label as evidence.
- No technical term without an explanation.
End with one line: "Self-check: all points met", or which point is open and why.
"done" means: the results right away with what's there. If there are no answers yet, send the three test questions and the test plan, and say that the count needs the pasted answers.

4. Follow-up instructions

Keep writing in the same test lead chat, so the AI already knows your questions and answers.

a) The same test in Perplexity Ask the same three questions in Perplexity, three times each in a new chat, and paste the answers with the labels P-Q1-R1 to P-Q3-R3. First, send this to the test lead:

I'm now asking the same three questions in Perplexity, three times each in a new chat. The answers will come with the labels P-Q1-R1 to P-Q3-R3. Count by exactly the same rules. Then one table: Question | ChatGPT (x of y) | Perplexity (x of y) | Named most often in both. Don't explain differences with guesses about how the two work. Only report what's in the answers.

b) What ChatGPT knows about you In a new temporary, unpersonalized chat, ask three times: "What do you know about [BUSINESS NAME] in [PLACE]? What do they offer, and who for?" Paste the three answers into the test lead chat with the labels K-R1 to K-R3, then:

Compare these three answers with my real details: [WHAT YOU OFFER, WHO FOR, SINCE WHEN, WHERE, HOW TO REACH YOU – BUSINESS DETAILS ONLY, NO CUSTOMER DATA]. Make a table: Detail in the answer | Label | correct / wrong / out of date / missing on my side. Add nothing from your own knowledge. For wrong details, say where I could correct them, and mark guesses "to be checked".

c) Check your website against the test questions

Here's the text of up to three pages of my website: [PAGE TEXT, NO CUSTOMER NAMES]. For each of my three test questions, check: does my website answer it in plain words? Answer: yes / partly / no, with the sentence from my website as evidence. Then: what's missing for a customer to find the answer on my site? Write text suggestions with [PLACEHOLDERS] for anything you don't know from my text (prices, figures, services). Don't invent anything about my business. Don't promise that ChatGPT will name me afterwards. Prices, reviews, advertising claims and professional rules: check with an expert.

d) Repeat the test and compare (for example in four weeks, in a new test lead chat)

Here are my last results from [DATE]: [PASTE RESULTS]. I've asked the same three questions again with the same setup; the new answers follow with the labels Q1-R1 to Q3-R3. Count by the same rules. Then a comparison table: Question | then (x of y) | now (x of y) | newly named | no longer named. Add: with three runs per question, a difference of one run is no evidence of a change. Don't explain differences with guesses.

e) Have a second AI check the count Open a new chat and paste this:

You're checking a count. Here are the answers I got in ChatGPT, each with a label: [THE NINE ANSWERS]. Here are the results another AI produced: [RESULTS]. My business is called [NAME AND SPELLINGS], my website is [WEBSITE].
Count it yourself: where am I named? Where does only my website appear as a source (that doesn't count as named)? Which other providers and websites appear, and how often?
Give me a list of every difference, with its label. If the results contain anything that's in none of the answers, list that first. Don't invent anything.

5. Example: what it looks like

MADE-UP EXAMPLE. The accounting practice, the town, every business and every website are invented. The nine "ChatGPT answers" aren't real either: we wrote them ourselves as test material, so we knew in advance what a correct count would be.

The filled-in block:

- My business name and other spellings: Ashdown Accounting, also "Ashdown Accountants" and "Ashdown Tax"
- My website: ashdown-anytown.com
- What I offer, in my customers' words: tax returns and bookkeeping for freelancers, the self-employed and new businesses
- Where my customers are: Anytown and the surrounding area
- Competitors I know of: [NAMES OR "NOT SURE"]
- Where I'm testing: ChatGPT, temporary chat, unpersonalized
- Date of the test: 15/10/2026

Two of the nine pasted answers (made up):

Q1-R2:
For freelancers in Anytown, these firms are worth a look:
- Bramley & Shaw Accountants: many freelance clients, good reviews.
- Calder Tax: transparent pricing, online meetings available.
Check whether the firm offers digital bookkeeping and how quickly they reply to questions.
Sources: anytown-business-directory.com · anytown-reviews.com · ashdown-anytown.com

Q1-R3:
Here are a few accountants in Anytown that could suit a freelancer:
1. Bramley & Shaw Accountants – lots of experience with start-ups and freelancers.
2. Ashdown Partners – advice for small businesses.
3. Pennington Bookkeeping – personal service, home visits possible.
Sources: anytown-gazette.com

In Q1-R2 the practice's website is among the sources, but its name isn't in the text: that's "source only". Q1-R3 has a similar name that isn't the practice. The test lead asked ("Is this your business under another name, or a different firm?") and only counted it as another provider after the "no".

Excerpt from the results. This is what came out of our test run with Claude on 15 October 2026. A second AI played the made-up owner, working from a profile. After three interview questions, the suggested test questions and nine pasted answers, these were the results. Shortened, otherwise unchanged. We knew the correct count in advance: named in 3 of 9, source only once. The results matched it exactly.

Snapshot from 15/10/2026, tested in ChatGPT (temporary chat, unpersonalized), 9 answers counted. The same question can bring different names tomorrow.

1. Your three test questions

Q1: I've just gone self-employed in Anytown and need someone to do my taxes. Who would you recommend?
Q2: Roughly what does a tax return cost for someone self-employed in Anytown, and who should I ask?
Q3: I'm self-employed near Anytown and want a fair price, to send receipts digitally and quick replies. What should I look for, and which accountants would suit me?

2. Every answer

Label	You named?	Position	Other providers in order	Sources
Q1-R1	yes	3	Bramley & Shaw Accountants, Calder Tax	3
Q1-R2	source only	–	Bramley & Shaw Accountants, Calder Tax	3
Q1-R3	no	–	Bramley & Shaw Accountants, Ashdown Partners, Pennington Bookkeeping	1
Q2-R1	no	–	no names	0
Q2-R2	no	–	Calder Tax, Bramley & Shaw Accountants	2
Q2-R3	no	–	Calder Tax	1
Q3-R1	yes	1	Bramley & Shaw Accountants	2
Q3-R2	yes	2	Calder Tax	1
Q3-R3	no	–	Bramley & Shaw Accountants, Pennington Bookkeeping	0

3. Result per question

Question	Runs	You named (x of y)	Rating	Named most often
Q1 Recommendation	3	1 of 3	Sometimes named	Bramley & Shaw Accountants (3 of 3)
Q2 Cost	3	0 of 3	Not named	Calder Tax (2 of 3)
Q3 Choosing	3	2 of 3	Mostly named	Bramley & Shaw Accountants and you (2 of 3 each)

Named overall: 3 of 9 answers. Source only: 1.

(Sections 4 and 5 shortened: Bramley & Shaw Accountants in 6 of 9 answers, Calder Tax in 5; most frequent source anytown-business-directory.com, in 4 answers.)

6. What to look at next

Cost question never names you. Q2-R1, Q2-R2 and Q2-R3 all had no Ashdown. Q2-R2 and Q2-R3 say Calder Tax "lists package prices on its website" and "shows its prices openly". Check whether your website shows prices or a clear way to get a quote, since customers ask "What would you charge for my tax return?"
One directory appears most often. anytown-business-directory.com is a source in Q1-R1, Q1-R2, Q2-R2 and Q3-R2. Check whether you're listed there, with current details.
Your website showed up, but your name didn't. In Q1-R2, ashdown-anytown.com was a source, yet the text named only Bramley & Shaw and Calder Tax. Check whether your homepage says plainly that you do tax returns for freelancers and the self-employed in Anytown.

(Points 4 and 5 and sections 7 and 8 shortened.) Last line: "Self-check: all points met".

What we changed afterwards: point 1 suggests putting prices on the website without "check with an expert". For an accounting practice, that's exactly the sensitive part. So that rule is now spelled out in section 6 of the instruction. In the second test run with the new version, the note came up next to prices.

6. How you'll know it's working

The snapshot line with the date is at the top, and nowhere does it say ChatGPT "always" answers this way.
Every name and every website in the tables appears in one of your pasted answers. Your spot check (step 13) matches.
"Source only" is listed separately and doesn't count as named.
Every point under "What to look at next" names the labels it's based on, and is something you can check yourself.

Common problems and what to do:

What happens	What you write or do
The test lead sends several questions at once.	"Just one question per message, please, as the instruction says."
ChatGPT names no businesses in an answer ("I can't recommend specific providers").	Not a problem. It counts as a run with "no names". Don't follow up, or it stops being a customer test. If it happens in all three runs of one question: "Reword Q2 so a customer is asking for providers", then ask the new question three times from scratch.
The answer shows no sources.	Not a problem. ChatGPT decides for itself whether to search the web. The test lead counts "Sources: 0".
You're named in run 1 and not in run 2.	Not a problem. That's exactly why you ask each question three times.
The test lead counts "source only" as named.	"In Q1-R2 only my website is among the sources. That's 'source only', not 'named'. Please recount."
The table has a name that's in none of the answers.	"Where does this name appear? Take it out."
The answers name businesses in another town.	Check that the town is in the question and that no VPN is running. Then ask again.
You ran a test in a normal chat by mistake.	Don't count that run; repeat it in a temporary, unpersonalized chat.
A temporary chat is gone before you copied it.	New temporary chat, same question, same label.
The conversation stops halfway.	Later, in the same chat, write: "Pick up where we left off." On the free plan, the reason is often the usage limit.

7. Limits and data protection

A snapshot, not a standing fact. Nine answers, on one day, from one place, with one setup. Other customers phrase things differently, sit somewhere else, have memory switched on, and may get other names. The test shows you what happened that day, not what always happens.
Three runs aren't many. A difference of one run between two tests can be chance. Only compare tests with the same questions and the same setup.
Nobody here knows how ChatGPT chooses providers. The test lead isn't allowed to claim it either. OpenAI itself says placement in search results isn't guaranteed, and that search results and citations can be incomplete, out of date or wrong. The points in section 6 of the results are things to check, not a recipe.
One thing OpenAI does spell out: for a website to be eligible for ChatGPT's search results at all, it has to let OpenAI's search crawler in (it's called "OAI-SearchBot"). Ask whoever looks after your website whether that's the case.
The test questions contain no personal data. They ask about providers, not people. Keep it that way. What doesn't belong in these chats:
- names, addresses, phone numbers, email addresses of customers or staff (not even as an example of "how a customer asked me last week")
- health data, bank details, ID numbers, salaries
- contracts, passwords, login details
- anything covered by professional confidentiality (accountants, doctors, lawyers, for example)
Business names in the answers are public information and are fine in the test lead chat. If your business carries your own name, that name is personal data about you. Whether you enter it is your call.
Don't use the results as advertising. Publishing "ChatGPT recommends us over X" is risky. Whether and how that's allowed is something to check with an expert. This isn't legal advice.
No bought or polished reviews, even if review sites keep showing up among the sources.

8. Next step

If you'd like to build processes like this with Claude yourself later on: our AI Blueprint is a ready-made setup for Claude, explained step by step, even if you don't code.

Go to the AI Blueprint