Most brands invest heavily in how their website looks. Very few know how they sound when an AI assistant answers a question about them. That gap matters more every month, because a growing share of people now ask ChatGPT, Perplexity, or Gemini before they ever open a search engine or a company page.
An AI detection audit is a simple discipline: you ask the assistants what they know about your brand, then compare their answers against reality. This post walks through how to run one yourself, what to look for, and how to think about fixing what you find, honestly.
Why an AI detection audit is different from an SEO check
SEO tells you where a page ranks. An AI answer is synthesized from many sources at once, and it often does not cite them. So the question is no longer "do I rank?" It is "does the assistant describe me accurately, and where is it pulling that description from?"
The answer can be wrong in three distinct ways:
- Missing — the assistant has no confident information and hedges or declines.
- Inaccurate — it states something false: wrong location, wrong offering, wrong founding year, or a competitor's detail attributed to you.
- Outdated — it repeats information that used to be true but no longer is.
Each failure mode needs a different fix, so the audit has to distinguish them rather than lump everything into "the AI got it wrong."
Step 1: Write your ground-truth sheet first
Before you touch any assistant, write down what is actually true. Keep it boring and factual:
- Legal or trading name, and common variations
- What you sell, in one plain sentence
- Where you operate
- Founding year, if relevant
- Contact channel
- Any fact people frequently get wrong
This sheet is your scoring key. Without it, an audit becomes vibes. With it, every AI answer becomes pass or fail on a specific claim.
Step 2: Ask the right kinds of questions
Run the same set of prompts across each assistant. Vary the framing, because assistants respond differently to direct versus discovery questions:
- Direct: "What is [brand name]?"
- Category discovery: "Who offers [your service] in [your area]?"
- Comparison: "How does [brand name] compare to alternatives?"
- Verification: "Is it true that [brand name] does [claim]?"
The discovery and comparison prompts are the honest test. It is easy to get a decent answer when you hand the assistant your exact name. The real question is whether you appear at all when someone describes the problem you solve without naming you.
Step 3: Record answers verbatim, then score against ground truth
Copy each answer exactly. Do not paraphrase, because the specific wording is the evidence. Then tag each claim in the answer:
| Claim in AI answer | Ground truth | Verdict |
|---|---|---|
| "Based in Jakarta" | Correct | Pass |
| "Founded 2015" | Actually 2019 | Inaccurate |
| "Offers logistics services" | Not offered | Inaccurate |
| (No mention in category query) | Should appear | Missing |
This table is the entire deliverable of an audit. It turns a fuzzy worry into a punch list.
Step 4: Trace where the answer likely came from
Assistants that show sources (Perplexity is the clearest here) make this easy. For those that do not, reason backward: a specific wrong founding year probably came from a directory, a data aggregator, or an old press mention. Search the exact phrase the assistant used and you will often land on the source that seeded it.
This matters because you rarely fix an AI answer by arguing with the AI. You fix it by correcting the underlying sources the model draws on.
Step 5: Fix the sources, not the symptom
This is where the honest work lives. If a third-party profile lists the wrong service, update it. If a fact is missing everywhere, it is missing because it was never published in a place the model trusts. The remedy is consistent, verifiable information across credible third-party platforms, not a single new page nobody links to.
This is exactly the problem The Resets Company works on. We help brands get found accurately by AI even when they do not run their own website, through three connected pieces of work: an AI Detection Audit to see what the assistants currently say, Distributed Brand Presence to build consistent, accurate information across credible third-party platforms, and AI Visibility Monitoring to track whether detection and accuracy actually improve over time.
We are deliberate about one thing: we separate what gets published from what the AI actually detects and verifies. Publishing information is an input, not a guarantee. We do not promise rankings or recommendations, because no honest provider can control what a model outputs. If you want to talk it through, reach us at hello@theresetscompany.com or read more at https://theresetscompany.com/.
Step 6: Re-audit on a schedule
One audit is a snapshot. Models update, sources change, and new errors creep in. Run the same prompt set monthly and keep your scoring tables. The trend line, more claims moving from Inaccurate or Missing to Pass, is the only measure of progress that means anything.
A note on honesty
Be skeptical of anyone who promises to make an assistant recommend you. What you can genuinely do is make the true information about your brand easy to find, consistent, and verifiable, so that when an assistant does describe you, it has accurate material to work from. That is the whole game: not manipulating the answer, but earning an accurate one.
Start small. Pick five prompts, run them across two assistants, and build your first scoring table this week. You cannot fix what you have never measured, and right now most brands have never once listened to how AI describes them.
Top comments (0)