How to run an AI visibility audit in an hour

An AI visibility audit is a one-hour check of five things: whether crawlers can read your site, what ChatGPT, Perplexity, Gemini and Google AI answer for your top ten buyer questions, which pages they cite, whether you're one of them, and the single fix to make first. Run the steps in order. Don't try all of it at once.

1. Confirm the crawlers can read your pages

Start here, because everything after it is wasted if the answer is no.

Open your robots.txt and read it line by line. Look for blanket disallows, and look for rules that name AI crawlers specifically. Then load one of your key pages with JavaScript disabled, or fetch the raw HTML, and check that the actual product description, pricing and comparison text is in the source rather than painted in by a client-side script after load.

Add llms.txt if you don't have one. Agency audits now treat llms.txt deployment, crawler governance and LLM readiness as part of the same technical pass as a technical SEO overview for extractability and citation readiness.

How to check it worked: fetch the page as plain HTML and search for your own product name and your pricing. If you can't find them in the raw text, a model probably can't either.

Common mistake: assuming a page that ranks in Google is therefore readable by every engine. Different crawlers, different rules, different rendering budgets. Check the raw file, not the rendered one in your browser.

Time budget: ten minutes.

2. Write ten buyer questions the way people actually ask them

Not keywords. Questions, in the words a buyer would type into a chat box at 11pm.

The standard approach is to start with your ten highest-value transactional keywords and write three to five natural-language versions of each. That's stage one of a four-stage audit that runs query identification, manual prompting, source mapping, then automated execution, and the whole point of stage one is mapping intent rather than matching keywords.

Group them across three intent types: informational queries, comparative prompts, and direct brand questions. Aim for a roughly even split, and include at least two prompts where a competitor is named alongside your brand.

So for an invoicing tool, you'd end up with things like "best invoicing software for freelancers", "is [you] or [competitor] better for recurring billing", "does [you] do multi-currency invoices".

How to check it worked: read your ten questions out loud. If any of them sounds like a search query rather than a sentence, rewrite it.

Common mistake: writing only questions you'd win. The uncomfortable ones are the audit.

Time budget: fifteen minutes.

3. Run each question on four engines and save every answer

Stage two is manual prompting and sentiment analysis, probing the engines directly. Open ChatGPT, Perplexity, Gemini and Google AI Mode side by side. Paste question one into each. Save the full answer text and every source link, not a summary.

Log four columns per answer: engine, prompt, brands named in order, sources cited.

One run per prompt is not enough, and this is the part most people skip. In welaunch's test of live AI answers (27 Sep 2026, variance section), only 9 of 32 brand appearances in ChatGPT answers (28%) were named in all four runs of the same prompt, and 6 of 21 brand appearances in Perplexity answers (29%) survived all four runs. Measured as set overlap, the brand lists ChatGPT returned for the same prompt across four runs matched at a mean Jaccard similarity of 0.54, and Perplexity at 0.57, per the same test.

That means a brand you see once might be a coin flip. Forbes puts it mildly: results from manual queries vary, but consistent patterns should emerge across prompters and query types. In an hour you can realistically run each prompt twice per engine. Do that, and flag anything that appeared only once as unconfirmed.

How to check it worked: every row in your sheet has a full answer saved, not a yes/no. You will want to re-read the wording later.

Common mistake: logging "mentioned: yes". Position matters, context matters, and whether the model called you "a cheaper option" matters more than the tick.

Time budget: twenty minutes, and this is the step worth stealing from the others.

4. Find the pages each engine is citing

Stage three is mapping source citations and tracking down the exact origins of what the models cite. Take every URL from step three and paste it into a second tab of your sheet. Count how often each domain appears.

The engines behave very differently here, so a result on one tells you little about the others.

How the four engines differ on sourcing, from welaunch's test of 434 live AI answers to buyer questions in seven software categories, 27 Sep 2026, US:

Perplexity hands you roughly three times the source list Google AI Mode does, and never came back empty in that test. Google AI Mode went sourceless in 21 of 111 answers. If you only audit one engine, you'll build a plan for a sourcing pattern the other three don't share.

Don't assume your Google rankings carry over either. Only 16.7% of sources cited in Google AI Overviews overlap with the first page of organic results, and AI Overviews now appear on 48% of all tracked search queries.

How to check it worked: sort your domain list by frequency. The top five domains are your actual target list.

Common mistake: collecting citations and never counting them. The count is the finding.

Time budget: ten minutes.

5. Check if Reddit, YouTube or a competitor is crowding you out

Now look at what's on that top-five list instead of you.

Across all 434 answers in welaunch's test (27 Sep 2026), 106 answers (24%) cited reddit.com. The weighting is lopsided by engine: Google AI Mode at 40% of answers, Perplexity at 31%, Gemini at 20%, ChatGPT at just 6%, per the same test. For buyer questions about choosing an AI visibility tool specifically, 22 of 80 answers cited reddit.com. For AI-visibility buyer questions, 12 of 63 answers cited reddit.com and 12 of 63 cited youtube.com.

So "get mentioned on Reddit" is a sensible plan if your buyers use Google AI Mode or Perplexity, and close to irrelevant if they live in ChatGPT. Your citation table from step four tells you which.

Then check which competitors keep showing up. In welaunch's test, Profound was named in 23 of 63 answers to AI-visibility buyer questions, with Otterly and Peec at 17 each. The pattern you want is not "who is biggest" but "which page about them is the engine reading".

How to check it worked: for each of your ten questions you can name the one page an engine would have to read to mention you.

Common mistake: treating a competitor's presence as a branding problem. It's usually a page problem. They're on a listicle, a directory, or a thread. You're not.

What counts as the one fix to make first

You will finish this audit with more things to fix than you can do at once. Pick one.

The question to ask: across your ten prompts, which single missing source appears most often? If a comparison roundup gets cited by three engines for six of your questions, getting named in that roundup is worth more than ten blog posts.

Three patterns, in the order they usually pay off:

• You're absent from a source cited repeatedly across engines. Fix: get onto it.
• You're named but described wrong. Fix: correct the page the model is reading, not your own homepage.
• You're nowhere, on anything, for every prompt. Fix: start with directories, because they're the fastest index to get into.

On that last one, welaunch's directory database lists 273 launch and listing directories, of which 198 accept a free listing. It breaks down into 163 SaaS and startup directories, 95 AI tool directories and 15 agent and developer directories, with 100 that accept AI products only. Submitting to the handful that already appear in your citation table beats submitting to all of them.

Do a free or low-cost version of this ai visibility check

The whole audit above costs nothing but an hour and a spreadsheet. If you want a second opinion, free analyses exist: Amplitude's free AI visibility analysis gives a visibility score and sentiment across top AI platforms including ChatGPT and Google AI Overview, plus competitive rankings, critical topics and related prompts and personalised recommendations for AI discoverability.

What a spreadsheet can't do is run itself. That's why the four-stage framework ends with moving from audit to automated execution, turning passive findings into continuous updates. Your sheet is a snapshot of one afternoon.

How often to repeat the audit

Given the variance numbers from step three, once a quarter is too slow and once a day is more than most founders will sustain. Weekly is about right for a small prompt set.

Two reasons. First, the same prompt returns different brand sets run to run: mean overlap of 0.54 on ChatGPT and 0.57 on Perplexity in welaunch's test, with only 28% and 29% of brand appearances holding across all four runs. A weekly run averages that noise out. Second, if you lean on a scoring tool instead, note that tool scores are not universal across tools and the metrics behind them keep changing. Compare a tool's score to its own previous score, never to another tool's.

What a realistic timeline looks like after you fix something

A new directory listing or a thread reply doesn't move an answer the same day. The honest expectation is that you're watching for a pattern, not an event, and consistent patterns emerge across prompters and query types rather than in a single check.

Practically: make one fix, keep running the same ten prompts weekly, and judge it on whether your name shows up in more runs per prompt over several weeks. Don't recheck an hour later and conclude it failed. Also don't swap your prompt list while you're measuring, or you've lost your baseline.

Turning a one-hour audit into a standing habit

The spreadsheet version works. It just decays, because nobody keeps rerunning the same prompts by hand every week.

If you want it to keep running, that's what welaunch does: it asks your buyer questions on ChatGPT, Perplexity, Gemini and Google AI several times a week, keeps every answer in full so you can read exactly what each engine said, finds the pages the engines cite, and drafts the reply, listing or page that gets you onto one of them. It works from your own accounts, you review every draft and press submit, and on Reddit or Quora it names the product with disclosure rather than stuffing links. The method is the same five steps. The difference is that it happens whether or not you remember.

FAQ

How do you measure AI visibility?

Count appearances across repeated runs, not single checks. Run the same ten buyer questions on each engine several times, log which brands are named and which sources are cited, then track how often your name appears per prompt over weeks. Core components of an audit include presence in AI responses, accuracy of positioning, and competitive visibility.

What's the difference between an AEO audit and a GEO audit?

In practice, very little. A GEO audit (generative engine optimisation) and an AEO audit (answer engine optimisation) both check what the engines say about you and which pages they're reading to say it. An LLM visibility audit is the same work under a third name. Pick whichever label your team uses and run the five steps.

Can ChatGPT do an SEO audit on itself?

You can ask it what it says about your category, and that's step three of this audit. What it can't do is tell you reliably whether it would name you, because the same prompt returns different brands across runs: in welaunch's 27 Sep 2026 test, only 28% of ChatGPT brand appearances held across all four runs of the same prompt. Run the prompt multiple times instead of asking the model to introspect.

What is AI visibility, in plain terms?

Whether an AI engine names your product when a buyer asks it a question, and whether it describes you correctly when it does. It depends heavily on which pages the engine cites, which is why source mapping matters more than your own site copy.

Do backlinks still help with AI search citations?

Less directly than in classic search. Forbes argues brands are no longer beholden to backlinks the way they are in traditional SEO, with AI platforms weighting contextual relevance, brand authority and consistency across sources more heavily. Being named on a page the engine already cites is the more direct route.

Can you pay to be cited in chat-based search results?

Not from the engines themselves. You can pay for placement on sources the engines read, which is a different thing, and some directories charge for a listing. Your citation table from step four tells you whether that specific source is worth paying for.

Sources

• AI Visibility Audit for Agencies — White Label IQ
• AI Visibility Audit: Step-by-Step | Yotpo
• How To Audit Your Brand’s AI Visibility
• welaunch.sh
• Get Your Free AI Visibility Report
• welaunch test of live AI answers, 27 Sep 2026
• welaunch's test of 434 live AI answers to buyer questions in seven software categories, 27 Sep 2026, US

Related on welaunch

• AI Overviews tracker comparison 2026: how to see whether Google's AI names you
• AEO vs GEO: the same job, two names, and where they actually differ
• AI search optimization: a 7-step guide for SaaS teams