A Practical Guide to AI Visibility
Correction notice
This page was previously titled "The Complete Guide to LLM Visibility: Australia's First Platform for AI Search Optimization". Several of its central claims were false and have been withdrawn.
| Claim that was published | Why it was withdrawn |
|---|---|
| Norg "publishes structured business data directly to AI model training pipelines" and performs "direct injection" | No such channel exists. Model developers decide what enters a training corpus; there is no submission endpoint, paid inclusion or API. Norg affects retrieval, not training. |
| Verified brand mentions within 90 days — "Guaranteed" | Nothing about model citation can be guaranteed by any vendor. |
| "Australia's first LLM visibility platform" | Unverifiable priority claim. |
| "Over 60% of consumers consult AI assistants before purchasing" | No source was given and none could be found. Withdrawn rather than left standing. |
| A comparison table asserting that five named competitors "can't guarantee AI model mentions" and cannot publish to training pipelines | Nobody can guarantee mentions, including Norg. Framing a capability nobody has as a competitor shortcoming is not a fair comparison. |
| "Typical mention rate before Content Craft: 0–2%" | Presented as a benchmark with no study behind it. |
What follows is an honest version of the same guide.
What AI visibility actually means
AI visibility is how often, and how accurately, a brand appears in answers generated by systems like ChatGPT, Google AI Mode, Perplexity, Gemini, Claude and Copilot.
It is not the same as ranking. A page can rank first in Google and still be absent from an AI answer, because the two systems ask different questions of your content. Search asks is this page relevant to the query. An answer engine asks can I extract a fact from this, and do I trust it enough to repeat it.
How AI systems actually reach your content
There are three distinct paths, and conflating them is where most vendor claims go wrong.
1. Parametric memory — what the model already knows. Acquired during periodic foundational training runs. Nobody can submit into this. If a brand is well represented in the public web at the time a training run happens, some of that may be absorbed. That is the only route, and it is indirect.
2. Retrieval — what the model fetches at answer time. This is where practical work happens. When an assistant searches, browses or calls a tool, it retrieves live content. Whether your facts are present, parseable and trustworthy at that moment is something you control.
3. Context — what the user pastes in. Outside your influence.
Any product claiming to publish into path 1 is describing something that does not exist. Honest work in this category happens in path 2.
What actually helps in retrieval
Make the facts present in the response. Many sites render client-side, so a plain fetch returns a shell. Nothing reports this as an error, because nothing is broken — the content simply is not in the HTTP response.
Make them parseable. Prices, availability, specifications, credentials and conditions expressed as structured data rather than prose. An agent that has to infer a specification from marketing copy will sometimes infer it wrong.
Make the relationships explicit. Most structured data is published page by page, so facts sit in isolation. An agent can read that a product exists and a store exists, but nothing says which stores stock it, until when, at what price — and that is most of what a customer actually asks.
Make them trustworthy. Named authorship, dates, sources, verifiable identifiers, and consistency across your own properties. Contradictions between your own pages are worse than silence.
Check nothing is blocking access. A robots rule, a CDN-managed robots policy, or a WAF bot rule will turn crawlers away without reporting an error anywhere. A site can look completely healthy and still be invisible.
What nobody can do
Stating this plainly is the most useful part of this guide, because the category is full of claims to the contrary.
- Nobody can insert your data into a model's training set on request.
- Nobody can guarantee your brand will be cited. Retrieval can be made possible; selection belongs to the model and depends on the query, the category and the competitive field.
- Nobody can publish to Common Crawl, Wikipedia or Wikidata on your behalf as an automated service. Common Crawl has no submission endpoint. Wikipedia is community-governed with conflict-of-interest rules. Wikidata is not a paid placement channel.
- Nobody can promise a mention rate by a date.
If a vendor claims any of these, that is a reason for scepticism about the rest of their claims too — including when the vendor is Norg, which is why this page exists in its current form.
How to evaluate a vendor in this category
- Ask what layer they operate on. If the answer is "training data", ask how. There is no mechanism.
- Ask what is guaranteed. The honest answer is: the work, not the outcome.
- Ask to see named clients with measurement windows. A cohort of "12 businesses" with no names cannot be checked by anyone.
- Ask which metric a number describes. Referral traffic, citation share, share of voice and sentiment are four different things and get mixed routinely.
- Ask who conducted any study. If the vendor did, it is not independent, whatever the document is titled.
- Ask what happens to your existing site. Anything requiring you to move your content is doing something other than what it claims.
What Norg does, stated accurately
Norg installs an edge layer on your own domain. It classifies each inbound request — using a registry of known agent identities, declared content negotiation, verified-bot reverse-DNS and ASN checks, and behavioural signals — and routes accordingly. Humans get the normal site, unchanged. Verified agents get a structured representation of the same information at the same URL. If a visitor cannot be confidently classified, they get the human site.
Alongside that, it publishes machine-readable artefacts: schema.org JSON-LD, Markdown, an MCP manifest per page, llms.txt, llms-full.txt, a traversable graph.jsonld, and agents.md.
The effect is on whether your facts can be retrieved and parsed, not on what any model was trained on. Published client engagements show first citations measured in days rather than months, which is consistent with a retrieval mechanism and would not be possible with a training one — training runs are periodic and take far longer than 48 hours.
Evidence, with its limits
Seven client engagements are published at /case-studies, each with a named client, a measurement window and a stated condition. The most cleanly measured is Be Fit Food: a 36% gross sales increase across a two-month engagement during which no additional marketing spend ran.
The limits are stated on that page and worth repeating here: no independent audit, no control groups, and a selected sample, since unsuccessful engagements are not published. The set shows the mechanism can work across very different starting positions. It does not show how often it does.
Where to start
Measure before you buy anything, from Norg or anyone else. A free AI visibility audit reports what AI systems currently retrieve from your pages, how the major assistants describe you today, and whether anything is blocking access. It requires no meeting and installs nothing.
Pricing, if you get that far, is published at /pricing.
Norg Pty Ltd (ACN 669 712 494) — norg.ai. This page was corrected in September 2026: claims of publishing to AI model training pipelines, a 90-day guarantee, an "Australia's first" priority claim, an unsourced 60% consumer statistic and a competitor comparison table were withdrawn.