The 20-Minute AI Visibility Self-Check

The short answer

Many buyers now ask AI assistants the questions they used to type into Google, and the assistants answer whether or not they know your business exists. The self-check below fits in about 20 minutes: run the twelve questions on the one assistant your buyers use most, or your six most important questions on all three of ChatGPT, Gemini and Perplexity.

The method has two passes, because they measure different things. Pass 1 asks each question plainly, the way a buyer would, and records whatever comes back, invented or not; that is what buyers are actually told. Pass 2 re-asks it in a fresh chat with an honesty guardrail attached, so the assistant admits what it does not know; that is what the model actually knows. The gap between the passes is the fabrication your buyers are hearing, and it points at which facts to fix at the source.

No tool required, no dashboard, no tricks at the end. The fixes that hold are an honest FAQ page and consistent facts, and the check tells you which to do first.

We make Baljia, an AI employee for founders and small businesses; it appears briefly near the end. The check itself needs only the assistants and a notes file.

How to run the check

• Use a fresh conversation for every ask, so one answer does not color the next.
• Use each assistant's default model, the one a buyer on the free tier gets.
• If the assistant offers to search the web, let it. That is what it would do for a buyer.
• Do not argue, correct or hint. You are measuring, not teaching. A correction in your own chat is a message in your chat, not an edit to the model.
• Write down the date. You will re-run this and compare.

Keep the scope tight so it stays a 20-minute job: all twelve questions on one assistant, or your six most important on all three. Run pass 1 for everything first. Then run pass 2 only for the questions where pass 1 named you, named a competitor, or stated facts; a plain "I do not know" in pass 1 needs no re-ask.

Replace the brackets in each template with your own words: [brand], [category], [city], [product], [competitor].

Pass 1: the twelve plain questions

Ask these exactly as a buyer would, with nothing added. You want the answer a stranger gets, including anything the assistant makes up.

1. The direct brand ask.

What is [brand]?

2. The what-and-who ask.

What does [brand] do, and who is it for?

3. The category recommendation ask.

What are the best [category] options for [audience]?

4. The local ask. Skip it if location is irrelevant to your buyers.

Recommend [category] in [city].

5. The comparison ask.

Compare [brand] and [competitor] for [buyer situation].

6. The "is it legit" ask.

Is [brand] legit? What do people say about it?

7. The buying-decision ask.

I need [product or service] to [specific job]. What should I buy and why?

8. The price ask.

How much does [brand] cost?

9. The alternatives ask.

What are the alternatives to [brand]?

10. The facts ask.

Who runs [brand], where is it based, and how long has it existed?

11. The problem-first ask. The question a buyer asks before they know your category exists.

How do I [the problem your product solves]?

12. The source probe. Ask this as a follow-up in the same chat, right after any answer that named you or a competitor.

Which sources did you use for that answer? List URLs if you can. If you
cannot cite any, say that plainly.

Pass 2: re-ask with the guardrail

In a fresh chat, repeat each question that earned a re-ask, with this attached:

[The same question.]

Rules for your answer: if you do not know something, say you do not know.
Do not guess or invent companies, people, prices or facts. Mark anything
you are unsure of as uncertain, and name real entries only; if you would
be padding a list, keep the list short instead.

Now read the two passes together. Where pass 1 stated a price, a feature or a founder and pass 2 says unknown, pass 1 was improvising, and buyers hear the improvised version. Where both agree and are wrong, a bad source is feeding the model, and the source probe from pass 1 is your lead. Where both agree and are right, that question is handled.

What to record for every answer

Five fields per answer, in a plain table or notes file. Resist recording feelings; record what was said.

Field	What to write down
Pass	Plain ask (pass 1) or guardrail re-ask (pass 2)
Named at all	Did your business appear, yes or no, and where in the answer
Facts right or wrong	Each fact about you, exactly as stated, marked right or wrong
Cited which URL	Every source it pointed to: your site, a directory, a list post, none
Named instead	Every competitor or alternative it recommended in your place

When the pass is done, let AI compile the notes, with the same honesty rules this whole check runs on:

Compile my AI visibility notes into a scorecard.

Rules:
- Use only the notes I paste. Do not add assistants I did not test,
  questions I did not ask, or results I did not record. If a cell has no
  note, write unknown.
- Count only unique real entries: one row per question per assistant per
  pass, and tell me the row count.
- No grades, no encouragement, no invented summary numbers. I want a
  record, not a pep talk.

My notes:
[Paste everything you recorded.]

Return: a table of question by assistant with named yes or no for each
pass; every wrong fact, worded exactly as the assistant said it; every
answer where pass 1 stated something pass 2 marked unknown; every URL
cited; and each competitor named instead of me, with a count of the
answers it appeared in.

What to fix first

First, wrong facts, starting with your own pages. If an assistant states your price, location or offering wrongly, check whether your own site says it plainly anywhere a crawler can read: no screenshots of pricing, no facts that live only inside an app. Wrong facts often trace to stale or third-party pages, and the only part you control is making the current facts easy to find on yours.

Second, an honest FAQ page. Take the questions from this check, plus the ones real buyers ask you, and answer them on one page in plain language, with the first sentence of each answer being the answer. This is the same page a human skims before trusting you, which is why it beats every trick: it is useful whether or not an assistant ever reads it.

Third, the cited URLs are your reading list. Whatever pages the assistants cited for your category, those are the places worth an honest presence: a directory entry with correct facts, a listing, a profile. Do not chase every list; chase the ones that actually got cited in your answers.

What not to do: tricks. Text written for crawlers instead of buyers, keyword-stuffed pages, instructions to AI hidden in your site. You would be optimizing against systems that change without notice, and the one asset that survives every change is a site whose facts are true, consistent and plainly stated. The full playbook for the honest route is in how to get your business recommended by ChatGPT.

Re-run the check monthly, same questions, and diff the notes. Answers shift as models and their sources update, so one pass is a snapshot, not a verdict.

Where Baljia fits, briefly: it is an AI employee for founders and small businesses. It writes and publishes SEO content that answers buyer questions, the honest FAQ page included, and it tracks Google Search performance and reports what happened. There is also a free audit at the SEO audit page, and plans are on pricing.

Frequently asked questions

How often should I run this check?

Monthly is enough, with the same questions each time so the comparison means something. Answers change as the assistants' models and sources update, on no schedule they publish, so treat any single run as a snapshot and judge the trend across runs.

Why do AI assistants get my prices and facts wrong?

Often because the answer was built from a stale page or a third-party list rather than your current site. You cannot edit the model, but you can keep one authoritative, crawlable version of every key fact on your own pages, so there is a correct source for the next update to find. The price ask and the source probe together show you where a wrong figure came from.

Can I just tell ChatGPT the correct facts about my business?

Not usefully. A correction inside your conversation changes that conversation, not what the assistant tells a stranger tomorrow. The durable fix is publishing the correct facts where the assistants' sources can read them, then re-checking next month.

What if none of the assistants knows my business at all?

That is the normal starting point for a small or new business, not a special failure. Start with the basics in order: correct facts on your own site, the honest FAQ page, then presence on the pages the assistants cited for your category. If you are pre-launch, fold this into launch week, since the directories and pages you set up there are exactly what gets cited later.

Why ask the questions twice instead of once with the guardrail?

Because the two versions answer different questions. The plain ask shows what a buyer is actually told, fabrications included; the guardrail re-ask shows what the model really knows when it is not allowed to improvise. Run only the guardrail version and you will never see the invented answer your buyers see; run only the plain version and you cannot tell a confident fact from a confident guess.

Run the first pass today

Twenty minutes, one assistant or three, five fields per answer. If you then want the FAQ page and the content written and published, and Search performance reported back to you, Baljia offers a 7-day free trial, no credit card, at https://baljia.ai/.

Sources

• https://baljia.ai/blog/how-to-get-recommended-by-chatgpt (checked November 5, 2026)
• https://baljia.ai/llms-full.txt (checked November 5, 2026)