What is an LLM Audit?

At a glance

• An LLM audit is a point-in-time read of how ChatGPT, Gemini, Claude and Perplexity describe your brand.
• It covers whether you appear, what gets said, and which sources the answer cites.
• Answers vary between runs of the same prompt, so one run is an anecdote, not a finding.
• The deliverable is a list of gaps and wrong claims, each traced to its source.

Reading back what the models already say

An audit starts with a prompt set: 20 to 50 questions a real buyer would ask about your category, your brand and your competitors.

Each prompt runs on each assistant you care about, more than once, because the same question does not return the same answer twice. Log the answers with their citations.

Then trace every claim back. A wrong pricing line usually comes from a stale page of yours. A competitor showing up in your comparison answers usually comes from a review site you have never edited.

What falls out is a short list: claims that are wrong, questions where you are absent, and sources you did not know were speaking for you. That list turns LLM visibility from a feeling into a work queue.

You cannot fix a story you have not read

Most brands have never read what an assistant says about them beyond a couple of curious searches. The gap between what the marketing site claims and what ChatGPT reports is usually wider than anyone expects.

Two failures come up repeatedly. The first is stale facts: old pricing, a discontinued product, a founder who left years ago. Models repeat whatever the web still says.

The second is absence. Nothing said about you is wrong, you simply do not come up. A shopper asks for the best option in your category and gets four competitors. Nothing on your site is broken. You were never in the set the assistant picked from.

Those two problems need different fixes, and reading the answers is the only way to tell them apart.

Audit your brand in an afternoon

No platform required for a first pass. Open a document and a browser.

1. Write 25 prompts across three types. Brand, category and comparison. Use the words a buyer would use, not your internal product names.
2. Run each prompt three times on two assistants. Fresh chats, no memory, no personalization. Paste every answer in with its citations.
3. Mark each answer three ways. Named, not named, or named wrongly. Highlight any claim you would not put on your own site.
4. Trace the worst five claims to a source. Click the citations. Most bad answers have one findable page behind them, often not yours.
5. Write the fix list and date it. Pages to update, third-party listings to correct, questions with no good page anywhere.

Re-running the same set is where the value sits. One audit tells you where you stand. Two tell you which way you are moving.