Most businesses that worry about AI visibility have never actually checked it. They have asked ChatGPT about themselves once, winced or celebrated, and moved on. That single query tells you almost nothing, because AI answers change from run to run, and because the reasons a business is missing from answers are usually invisible from the answer itself.
This is the audit we run before any engagement, published in full. It takes about two hours with no paid tools, it produces a written list of specific problems, and every check comes with a pass mark so you know where you stand. The thresholds come from our own measured studies of Australian businesses, not from theory.
Part one: can AI reach you at all
Check 1. The crawler access test
This is first because when it fails, nothing else matters, and it fails silently more often than anything else in this field. When we tested hundreds of Australian business sites in August 2026, 13 per cent served a normal page to a browser and an error to an AI crawler user agent. The culprit is almost never a deliberate choice. It is a firewall or CDN bot protection setting, often a default, quietly returning 403 or 429 responses to GPTBot, ClaudeBot and PerplexityBot while every human sees a working site. We have found this on marketing agency websites, including ones selling AI SEO.
How to check: use a command line tool or an online user agent tester to request your homepage with the user agent set to GPTBot, then ClaudeBot, then PerplexityBot, and compare the response code to a normal browser request. Anything other than 200 across the board is a fail. The fix is usually one rule in your firewall or CDN dashboard allowing verified AI crawlers.
Check 2. robots.txt
Read your robots.txt file and look for AI crawler names. Our crawl found 22 per cent of Australian sites name at least one AI crawler and 10 per cent fully block one. Blocking training bots is a legitimate choice, but blocking search-facing crawlers such as OAI-SearchBot or PerplexityBot removes you from AI answers on purpose. Decide deliberately. Pass mark: no search-facing AI crawler blocked unless you chose it knowingly.
Check 3. Is your content readable without JavaScript
Fetch your key pages with JavaScript disabled and check whether the actual content appears. Content that only exists after client side rendering, inside click to open accordions, in PDFs or baked into images is invisible or unreliable for most AI retrieval. Pass mark: your services, prices, credentials and contact details all present in the raw HTML.
Part two: is there anything worth quoting
Check 4. The six facts test
When an assistant recommends a business, its stated reason is nearly always a fact quoted from that business’s own site. We audited 500 Australian homepages for the six facts assistants reach for: service area, pricing or a pricing approach, licence or ABN, year founded, response time, and review presence. The average homepage carried 1.94 of the six. Twenty two per cent carried none. Only 3 per cent carried all six. The full numbers are in that study.
Score yourself honestly against the list of six. Each one should be in plain text on your homepage or one click from it. Pass mark: four or more. Publishing all six puts you in the top 3 per cent of Australian businesses on this measure alone.
Check 5. The register test
If your industry has a public register, licence board or accreditation scheme, assistants check it by name. When we studied solar recommendations, ChatGPT ran literal lookups against accreditation registers using site: operators before recommending installers, and a third of its background searches targeted a register or industry body. The same behaviour shows up for builders, lawyers and brokers. Search the register yourself for your exact trading name. If your listing is missing, outdated or under a different name than your website uses, an assistant verifying you will fail the match and drop you silently. Pass mark: your register entry exists and matches your website’s trading name exactly.
Check 6. Structured data
Run your homepage through a schema validator. You are looking for LocalBusiness or your specific subtype rather than a bare Organization block, with address, hours, service area and, where legitimate, aggregate rating filled in. Our homepage study found JSON-LD on 81 per cent of sites but the useful business types on only 37 per cent, and FAQPage markup on just 16 per cent. Implementation details are in our schema guide. Pass mark: correct business subtype with real data, plus FAQPage where you genuinely answer questions.
Check 7. llms.txt
Check whether yourdomain.com.au/llms.txt exists. About 31 per cent of Australian domains now have one, but a third of those were auto generated by an SEO plugin and 32 per cent contain none of the six quotable facts. An auto generated file is a sitemap wearing a costume. Ours is a short, hand written summary of who you are, what you do, where, for how much, and what proof exists. Pass mark: a hand written file carrying at least four of the six facts. The adoption data is in our llms.txt study.
Part three: what the assistants actually say
Check 8. The prompt panel
Now, and only now, ask the assistants. Write down ten questions a real buyer would ask, using their words, not yours. Best providers in your city, cost questions, is X worth it questions, and your brand name directly. Run each through ChatGPT with web search, Perplexity, and Google’s AI Mode. Record three things for each: were you named, was your site cited as a source, and who was named instead.
Run the panel twice on different days before concluding anything. We measured citation overlap of only 24 per cent between identical back to back runs, so a single run is noise. What you are establishing is a baseline rate: named in three of ten prompts, cited in one, whatever it is. That number is what improvement gets measured against. The behaviour differences between engines are covered in our AI Mode study and our volatility study.
Check 9. The fan out test
When ChatGPT researches a recommendation, it runs background searches, typically in the pattern of brand plus service plus city plus official. Search Google yourself for your business name plus your main service plus your city. If that search does not return a page of yours that confirms the combination, you fail the exact lookup assistants run to verify you. This is one of the most common silent failures we find: the business is real, the service is real, but no single page says both things together. Pass mark: one page that states your name, your core service and your city in plain text.
Check 10. The third party citation map
Look at who was named instead of you in check 8, then look at which sources the assistants cited. In most Australian categories, a handful of directories, review platforms and listicles write the shortlists. In residential solar it is comparison sites in 14 of 16 answers. In trades it is directories and Reddit. Make the list of the five sources that keep appearing in your category, then check whether you exist on each. Pass mark: present and accurate on at least three of the five sources AI cites in your market.
Check 11. Reviews where they count
Assistants quote review counts and ratings next to business names, and they read them from Google profiles and the platforms in check 10. Note your review count against the businesses that got named. If they hold hundreds and you hold twelve, that gap is part of why they are named. The fix is a routine, not a campaign, and we covered the mechanics in our review generation guide.
What the audit looks like on a real business
An example makes the sequence concrete. A Melbourne trade business we audited this year, established, well reviewed, profitable, scored four of eleven. Reachable and indexed, fine. Then the failures stacked up in a pattern we see constantly. Their firewall was returning errors to GPTBot, switched on by their hosting provider during a security update nobody remembered. Their homepage carried one of the six facts, the service area, with pricing, year founded and licence living in a PDF capability statement. Their register entry used the directors’ company name while the website used the trading name, so the exact lookup an assistant runs would fail. And their prompt panel baseline was two named mentions across sixty runs, both on Perplexity, neither citing their site.
The repair order followed the audit order. Week one: the firewall rule, ten minutes once the right person was found, and the facts moved from PDF to homepage text. Week two: the register alignment and schema. Month two onward: the citation layer, which for their category meant two directories and a review platform doing most of the naming. By the third monthly panel run, citation rate had moved from 3 per cent to 14 per cent of runs, led by Perplexity exactly as the engine behaviour predicts, and naming followed the quarter after, at which point the conversation changed from whether AI visibility was real to which services deserved pages next.
The point of the example is not the specific numbers, which vary by category and city. It is that nothing in the repair required special access or expensive tools. Every fix was available to the business the whole time. What the audit supplied was the order, and the order is most of the method: access before facts, facts before citations, citations before content volume, measurement wrapped around all of it.
Scoring and what to do next
Count your passes out of eleven. Most Australian businesses we audit score three or four. Seven or more puts you ahead of nearly every competitor you have. The order of repair matters more than the score: fix access first (checks 1 to 3), because nothing downstream works without it. Then publish facts (checks 4 to 7), because that is what answers are built from. Then work the external layer (checks 9 to 11), because that is slower and compounds. Re-run the prompt panel monthly and track the rate, not the individual answers.
If you would rather have it done for you, this exact audit is the first deliverable in every plan we offer, with pricing published in our cost guide. But the two hour version above is genuinely free and finds the worst problems, and a business that runs it will brief any agency far better than one that has not.
Frequently asked questions
How often should I audit AI visibility?
The full audit twice a year. The prompt panel monthly, because that is your measurement layer. Crawler access whenever you change hosting, CDN or security settings, because that is when it silently breaks.
What is the most common failure?
Two tie. Firewalls silently blocking AI crawlers, which we found on 13 per cent of reachable Australian sites, and homepages publishing almost none of the six facts assistants quote, where the national average is under two of six. Both are cheap to fix and most businesses have never checked either.
Do I need paid tools for this audit?
No. Every check above uses free methods: user agent testing, schema validators, the assistants themselves and your own search. Paid monitoring tools earn their keep later, at the tracking stage, once there is a baseline worth tracking against.
My business appears in some answers already. Should I still audit?
Yes, because appearance is unstable. Identical queries produced only 24 per cent citation overlap in our testing, so a business that shows up today on thin foundations disappears tomorrow without anyone choosing it. The audit tells you whether your visibility is anchored to facts and citations or riding on luck.
How is this different from a normal SEO audit?
It overlaps on technical foundations but measures a different endpoint. An SEO audit asks whether you can rank in a list. This audit asks whether an assistant can reach you, verify you, quote you and name you in a generated answer. Passing one does not mean passing the other, which is exactly what GEO vs AEO vs SEO covers.
What does a professional AI SEO audit cost?
As a standalone piece, Australian agencies charge a few hundred dollars to around two thousand depending on depth. Most fold it into the first month of a retainer. Run the free version first either way, because you will know within two hours whether you have the two killer problems, and you will negotiate from knowledge rather than fear.
