How to run an AI visibility audit that tells you something.
Most first audits produce a number and no decision. Pick questions a buyer would really type, check the answer engines that matter, and stop one variable answer from turning into a confident wrong conclusion.
· By the Askwords team
An AI visibility audit is easy to do badly. Ask a model whether it knows your brand, watch it say yes, and conclude that you are fine. You have learned nothing, except that a model is agreeable.
A useful audit answers a harder question: when someone who has never heard of you asks for a recommendation, does the answer include you, and if not, what did the named brands have that you did not? Budget about ninety minutes for the first one. It goes like this.
A baseline you can argue with, not a score you can post
The output of a good audit is not a percentage. It is a short list of sentences like “on four of our six buying questions, ChatGPT recommends three competitors and cites a G2 category page we are not on.” That is a finding. Somebody can act on it on Monday.
A percentage on its own cannot be acted on, and worse, it invites the wrong reflex: chasing the number instead of the cause. Treat the score as a way to notice change over time, and treat the answer text as the thing you actually read.
Everything downstream depends on getting this right
People rush this part, and it decides whether the audit is worth anything. Five well-chosen questions beat a hundred generated ones, because a hundred generated prompts mostly measure how many ways there are to phrase the same thing.
Pull them from real sources: your sales calls, your support inbox, the first line of your inbound emails, the questions your best customer asked before signing. Then spread them across four tiers.
| Tier | Example | Why it earns a slot | Mix |
|---|---|---|---|
| Category | “best project management tool for a 5-person agency” | The unbranded question a buyer asks before they know you exist. This is where absence costs you money. | Use 2 to 3 |
| Comparison | “Notion vs Asana for a small marketing team” | Shows whether you are framed as a real alternative or a footnote. Include one where you are the underdog. | Use 1 to 2 |
| Problem | “how do I stop client work slipping through the cracks” | The pain-first phrasing that never appears in your keyword tool. Models answer these with brands anyway. | Use 1 to 2 |
| Brand | “is Acme any good” | Reputation control, not demand capture. Useful, but it is the last question to add, not the first. | Use 1 |
Questions that teach you something
- Phrased the way a person speaks, not a keyword tool
- Contain a constraint: team size, budget, industry, city
- Could plausibly be answered without naming you
- Map to revenue you actually want
Questions that flatter you
- “What is [your brand]?”: it will say something nice
- One-word categories with no buyer in them
- Questions only your existing customers would ask
- Fifty variants of the same phrasing
A note on demand
Check at least three, and know what each one is for
The same question returns different brands on different answer engines. An audit of one answer engine is an audit of one product’s opinion. Auditing Google AI Overviews and ChatGPT alone covers the majority of the answers your buyers will actually see; adding Google AI Mode, Gemini and Perplexity fills in the rest.
Weight them by reach, not by convenience. Perplexity is the most generous citer in the group and therefore the most tempting to audit. It also has the smallest audience of the five. A win there is a useful clue about what evidence works. It is not the same as a win on AI Overviews.
Sign out first
Your logged-in ChatGPT knows your company, has your past chats in memory, and will happily recommend you. That is not the answer a stranger gets. Audit in a fresh incognito window with no account, or through a tool that requests answers without personal history. The gap between the two is often the whole finding.
Ask twice before you believe anything
One number should change how you audit. Paste the same question twice in a row and the brands you get back typically overlap by only about a third to a half. No competitor published anything in the minute between your two runs. The machinery underneath simply moves.
Which means a single run is a sample, not a measurement. If you audit once, find yourself absent, and rebuild your content strategy around it, you may have redesigned your quarter around a coin flip.
Repeat each question
Three runs per question per answer engine, then take the majority result. Two out of three beats one out of one.
Keep the wording frozen
Re-audit with the exact same strings. A reworded question is a new question, and its result cannot be compared to the old one.
Re-check across days
Treat a change as real only once it survives a second look on a different day. One day’s opinion is one observation.
Four fields per answer, no more
A spreadsheet with one row per question per answer engine is enough for a first audit. Resist adding columns you will never fill in twice.
- 1
Position
Cited, mentioned, buried or absent. One word, per question, per answer engine. This is the only field you will trend.
- 2
The sources
Every URL the answer linked. Count how many are yours, how many are third-party pages about you, and how many belong to a competitor.
- 3
The slate
Which brands were named, in order. Position one in an AI answer is worth more than position one in Google, because there is no position eleven.
- 4
The framing
The clause the model used about you. “Good for beginners” and “enterprise-grade” send very different buyers, and only one of them is yours.
The fourth field is the one people skip and the one that most often changes a roadmap. Being described accurately is a separate battle from being mentioned, and you can lose it while winning the first.
When the spreadsheet stops scaling
Turn a full spreadsheet into one job
Sort your rows by position, worst first, and look only at the absent ones on your category questions. Then open the sources those answers cited. One of five things will be true, and each points somewhere different:
- The cited pages answer a question you have never answereda content gap
- The cited pages are third-party lists you are missing froma mentions and PR gap
- The model describes you wrongly or vaguelyan entity and clarity gap
- Competitors have reviews and ratings you do nota proof gap
- Your pages are excellent and still never citedcheck crawler access before anything else
Pick the one that appears most often across your absent rows. Do that one. Re-run the same audit in four to six weeks, long enough for pages to be crawled and re-read, short enough that you still remember why you did it.
A good audit is repeatable, boring and specific.
Keep reading
Sources and further reading
- Google Search Central: Optimizing your website for generative AI features
- Google Search Central: AI features and your website
- Ahrefs: The complete AI visibility guide
Askwords measures a selected set of buyer questions against fresh answers from each engine. AI answers vary by model, time, context and personalization, so a check is evidence of the answers observed, not a universal judgement about a brand.