Short answer: write six questions the way a customer who does not know you yet would write them, without your company name. Run them on two AI platforms in a clean chat. Save every answer with the date, the exact question, the brands mentioned in order and the sources. Repeat a week later. If you appear in 0 of 12 answers, you have a discovery problem. If you appear with wrong facts, a representation problem. If competitors appear citing sources you are not in, a sources problem. Each one is worked differently.
Why "do you know My Company Ltd?" is the wrong question
Put your name in the question and almost any model will answer something. That measures nothing except the model's ability to complete a sentence. What matters is what happens when a buyer asks about your category without knowing you. That is where the shortlist is decided.
Step 1. Six questions, written like a customer
Two discovery, two comparison, two decision. For an online accounting firm in Spain:
- Which online accountant would you recommend for a freelancer in Madrid?
- Best online accounting firms for a small limited company
- Online accountant versus a local firm: what is the difference?
- Alternatives to the big online accounting platforms for a small business
- How much does an online accountant cost for a freelancer?
- Is it safe to file VAT through an online accountant?
Write them as questions, not keyword strings. If you have access to the questions people ask your sales or support team, use those. They beat anything you invent.
Step 2. Two platforms and a clean chat
Pick two where your customers actually ask. For most businesses, ChatGPT and Claude are a good starting pair; Gemini is the usual third, and Google AI Overviews depends heavily on the query and the location.
Use a temporary chat or a session with no history. ChatGPT can use memory and previous conversations to personalise answers, which contaminates the measurement. If you usually ask about your own business, the model knows you, not your market.
Step 3. Save every answer as if you had to defend it
A screenshot is good for showing. It is useless for comparing. For each question and platform, save:
| Field | Example |
|---|---|
| Date and time | 11 Sep 2026, 10:40 |
| Platform | ChatGPT (web search on) |
| Exact question | Which online accountant would you recommend for a freelancer in Madrid? |
| Full answer | The whole text, not a summary |
| Brands mentioned, in order | Brand A, Brand B, Brand C |
| Do you appear? | No |
| Sources or domains cited | domain1.es, domain2.com |
| Claims about your brand | None |
Twelve rows. Twenty minutes. You now have a baseline you can repeat.
Step 4. Repeat in a week
Answers change: with the date, the platform, small changes in wording. One round tells you what happened that day. Two rounds tell you whether it was stable or a coincidence. Label each row stable (same mention, similar sources) or variable (the mention, the order or the main source changed).
Step 5. Read the result
You appear in no answer at all. The most common outcome for small businesses, and a discovery problem. First check the basics: your site is indexed, you are not blocking AI search crawlers in robots.txt, and your home page says in one sentence what you do, for whom and where. Then look at which pages are cited in the answers where others do appear: usually comparisons, directories, pricing pages and FAQs. Those are probably the pages you are missing.
You appear, but with wrong facts. A representation problem. Find where the error comes from (an old listing, an outdated directory, an ambiguous page of your own) and fix what you control first: your website, your official profiles and your organisation's structured data. What you do not control, you request, politely, with the correct source linked.
You appear rarely and competitors cite sources you are not in. A sources problem. There is no honest shortcut: either you create the page that answers that question better than anyone, or you get into the external source with a real identity (a legitimate directory, a publication, a community where you take part under your own name). Buying reviews or mentions is not a route; it is a risk.
What this method does not tell you
- It does not give you market share. "I appear in 4 of 12 answers" describes a sample of twelve, not what every buyer sees.
- It does not tell you why. A source cited next to an answer does not prove it caused it.
- It does not predict tomorrow. That is why you repeat it.
Mistakes we see often
- Asking with the brand name and calling it a day.
- Doing it once and drawing conclusions.
- Using an account with history and memory enabled.
- Confusing "it cited this site" with "this site is the cause".
- Firing off a hundred questions with no criteria instead of twelve with buying intent.
What to do next
If you have your twelve rows, you already know more than most of your competitors. The next step depends on what you saw: a pricing page, an FAQ that actually answers, a data correction, or getting into the source that already cites others.
If you would rather we did it, AnswerGravity's free review looks at three questions in your category on two platforms and replies by email with what we observed and where we would start. The full diagnosis is twelve questions, twenty-four saved observations, up to three competitors and five priorities, for €149 plus VAT.