An AI visibility audit answers a simple question: when someone in your category asks an AI assistant what to buy, does the answer name you? Doing it properly is mostly about discipline: fair questions, enough asks, and honest counting. Here’s how to run a basic version yourself.
What an audit tells you
A useful audit answers four things:
- Who gets named. Every brand that appears in the answers, and how many answers named each one.
- Where the answers look. The websites cited, especially in answers that left you out.
- What they say about you. Whether you’re listed as a top pick, mentioned with a note of caution, or described in a way that doesn’t match your own website.
- What changed. On a second run with the same questions, what moved and what didn’t.
Step 1: write fair questions
The questions decide everything, so take care here.
- Write as a buyer, not a marketer. “Which term plan is best for a 30-year-old in India?” not “best term insurance solutions”.
- Cover the stages of buying: first looking (“what should I look for in a mattress?”), comparing (“memory foam or latex for back pain?”), choosing (“which mattress brand is best under ₹30,000?”).
- Cover each product line you sell. A brand strong in one line can be absent in another.
- Add the market where it changes the answer, for example “in India”.
- Don’t name yourself. A question that contains your brand name will almost always mention you, and measures nothing.
- Test each question once. If an answer to it names no brands at all, it can’t tell you anything about who gets named. Rewrite or drop it.
Ten to twenty-five questions is a reasonable start.
Step 2: ask each question more than once
ChatGPT gives a different answer each time you ask. One answer per question is a coin toss; three tells you whether your position is settled.
- Ask each question three times.
- Start a new chat each time, so earlier conversation doesn’t shape the next answer.
- Use the same settings throughout. Note whether the answer used web search. Searched answers show inline citations and a sources list, according to OpenAI.
- Note whether each answer searched. In an Ahrefs study of 66.6 million responses from free, logged-out ChatGPT (August–September 2026), about one in seven contained fan-out queries, the shorter searches ChatGPT writes for itself. Keep searched and unsearched answers in separate counts. An answer without search reflects what the model learned in training, which stops at a training cutoff.
With 20 questions asked three times, you’ll have 60 answers. That’s a real sample.
Step 3: record everything
Copy each answer into a spreadsheet in full, word for word. Don’t summarise. You’ll want to re-read them later.
| Column | What goes in it |
|---|---|
| Question | The exact words asked |
| Ask number | 1, 2 or 3 |
| Date | When you asked |
| Web search used? | Yes or no |
| Full answer | The complete text |
| Brands named | Every brand, as written |
| Websites cited | Every source link |
| Named us? | Yes or no |
Watch the spelling of brand names. An answer might write your brand three different ways, so search for every version your brand goes by before marking “no”.
Step 4: count honestly
This is where most audits go wrong.
- Use counts, with their total. A line like “named in 18 of 60 answers” tells you something. A percentage on its own hides how many answers it came from, and a “score out of 100” hides how it was calculated.
- Count every brand the same way. Then sort the list by how many answers named each one. That list shows who ChatGPT treats as the options in your category.
- Describe your search, not the world. If you weren’t named, the honest sentence is “not named in any of these 60 answers”. That describes your search. It doesn’t prove you’re absent from every answer ChatGPT gives.
- Keep searched and unsearched answers apart. They measure different things, so don’t add them together.
Step 5: look where the answers looked
Now filter to the answers that didn’t name you, and list the websites they cited. Count how many of those answers cited each site.
This list is often the most useful output of the whole audit. It might be your competitors’ own product pages, a comparison site, a marketplace or a regulator. Each one is a concrete place to look: is your brand on that page, and is what it says about you current?
Also check the answers that did name you. Did they cite your own website, or did they learn about you somewhere else?
Step 6: run it again later
Act on what you found, then ask the same questions the same way a few weeks later. Comparing two runs shows what changed.
Be careful with what you conclude. Answers vary on their own from one ask to the next, so a small change can be noise. And two runs show a difference, not proof that your work caused it. Never compare two runs that asked different questions.
Where a do-it-yourself audit gets hard
The method is simple. Doing it well at scale is not:
- Volume. Sixty to seventy-five answers, each copied in full and read for every brand and link, is a long afternoon, and it has to be repeated the same way each time.
- Name matching. Brands appear under short names, old names and misspellings. Missing one variant makes a count wrong.
- Fair questions. It is easy to write questions that flatter your own brand without meaning to.
- App versus model. The ChatGPT app adds memory and its own settings, so answers can differ between people and between the app and the underlying model.
How Citationly helps
Citationly runs this audit for you: questions prepared for your category, each asked three times on ChatGPT with web search on, every answer recorded word for word, and every count opening to the answers behind it. The AI Visibility Audit is ₹10,000 + taxes and includes a second run on the same questions 4 weeks later. See pricing or book a demo.