Generative Engine Optimisation: What It Is and What It Isn't

Correction notice

This page was published under the title "Suggestion 4". That was the identifier of an unreviewed content suggestion, not a title. It went live as a public document with that name, which is a publishing failure in itself.

The content also carried claims that have been withdrawn:

Claim that was published Why it was withdrawn
GEO "publishes directly to the models themselves" and establishes "direct data relationships with AI systems" No such channel exists. Model developers control training corpora and there is no submission endpoint, partnership tier or API for third-party data.
Named-competitor criticism of Surfer SEO, Semrush, Ahrefs and Frase.io Those are real products built for a real job. Norg cannot substantiate claims about their internals, and criticising them for lacking a capability nobody has is not a fair comparison.
"AI-sourced traffic converts at significantly higher rates" from "early data" No study, no sample, no method.
Norg "specialises in" financial services, insurance and legal Norg has no published client in any of those sectors.

What follows is an accurate account of the same subject.


What GEO actually is

Generative Engine Optimisation is the practice of making a brand's facts available, parseable and trustworthy at the moment an AI system generates an answer.

That is a narrower claim than the term usually implies, and the narrowness is the point. GEO is not a way to influence what a model was trained on. It is work on the retrieval path — the moment an assistant searches, browses or calls a tool and has to decide what to say about you.


The three layers, and which one is workable

Parametric memory. What the model absorbed during training. Periodic, controlled entirely by the model developer, with no submission route. A brand widely represented on the public web when a training run happens may end up better represented in the model — but that is a consequence of being genuinely prominent, not a service anyone can sell.

Retrieval. What the model fetches at answer time. This is the workable layer, and everything credible in this category happens here.

Context. What the user provides in the conversation. Outside anyone's control.

When a vendor claims to work on layer one, ask for the mechanism. There isn't one.


Why GEO is not simply SEO

The two overlap more than vendors like to admit, but they optimise for different questions.

Search asks: is this page relevant to this query, and how does it compare to other pages?

An answer engine asks: can I extract a specific fact from this, and do I trust it enough to state it in my own voice?

That difference has consequences. A page can rank first and still be useless to an answer engine, because the fact the agent needs is embedded in prose, or in an image, or arrives only after JavaScript execution. Conversely, a modest page with clean structured data can be cited repeatedly.

The practical divergences:

None of that makes SEO obsolete. Content that ranks well is usually content that retrieval systems also encounter.


What the legacy tools actually do

The withdrawn version of this page attacked several named products. A fairer account:

Tools like Surfer SEO, Semrush, Ahrefs and Frase.io are built to analyse and improve performance in search engines. They do that job, and they do it against a well-understood target. They were not designed to report whether Claude mentions your brand, because that was not a question anyone was asking when they were built.

That is a difference in scope, not a defect. Several have since added AI-visibility features of their own. Anyone evaluating this category should compare current products on current capabilities, rather than trusting a competitor's characterisation — including this one.


What actually works

  1. Check what a plain fetch returns. Not what the browser shows. If the HTTP response is a shell, nothing else matters yet.
  2. Check nothing is blocking access. Robots rules, CDN-managed robots policies and WAF bot rules all turn crawlers away without producing an error anywhere. This is the single most common cause of total invisibility, and the cheapest to fix.
  3. Express facts as data. Prices, availability, specifications, credentials, conditions — as structured fields, not sentences.
  4. Connect the entities. Publish relationships, not just isolated records.
  5. Make claims checkable. Named authors, dates, identifiers, sources. Keep your own pages consistent with each other; internal contradictions are worse than silence.
  6. Keep it current. Stale structured data is worse than none, because it is confidently wrong.
  7. Measure. Track which systems retrieve your content, how often, and what they say about you.

None of that requires a vendor. A competent team can do all of it. A platform makes it faster and keeps it consistent at scale, which is what Norg sells — not access to anything closed.


Honest limits of the category

Nothing can be guaranteed. Retrieval can be made possible; selection belongs to the model.

Measurement is immature. Citation share depends entirely on which queries you measure against, and there is no industry-standard query set. Two vendors can report different numbers for the same brand and both be reporting honestly.

Attribution is hard. AI-referred traffic is under-counted by tools that filter agent visits as bot noise, and over-claimed by vendors with an incentive to count generously.

Nobody has longitudinal data. The category is a few years old. Claims about what compounds over a decade are speculation.


Where Norg fits, stated plainly

Norg installs an edge layer on your own domain that classifies requests and serves verified agents a structured representation of the same information humans see, at the same URL, alongside machine-readable artefacts — schema.org JSON-LD, Markdown, an MCP manifest, llms.txt, a traversable entity graph.

It operates on retrieval. It does not touch training. It cannot guarantee citation. Norg also has a commercial interest in you concluding this work matters, which is worth holding in mind while reading anything on this site.

Seven client engagements with named clients, measurement windows and stated conditions are at /case-studies, along with what they do not demonstrate. The free audit measures your own position without a meeting or an install.


Norg Pty Ltd (ACN 669 712 494) — norg.ai. This page was corrected in September 2026: it was retitled from the internal identifier "Suggestion 4", and a direct-to-model mechanism claim, named-competitor criticism, an unsourced conversion claim and a regulated-industry specialisation claim were withdrawn.

What was wrong with the earlier version of this page?

It was published from an unreviewed generator suggestion titled 'Suggestion 4' with an empty description, and it asserted that Norg 'publishes structured data directly to the large language models themselves'. That is not possible for any vendor. It also criticised four named competitors — Surfer SEO, Semrush, Ahrefs and Frase.io — against that non-existent capability.

What is generative engine optimisation?

The practice of making a business's facts available, parseable and trustworthy at the moment an AI system generates an answer. The unit of work is the fact — a price, a location, an opening hour, a specification — published in a machine-readable form on the business's own domain, where a retrieval system can find it.

What is it not?

It is not submitting data to model developers. It is not paid inclusion in an AI answer. It is not a ranking system with positions you can buy or climb. It is not keyword optimisation with a new label. And it is not a guarantee that any model will cite you.

Can anyone publish into a large language model?

No. Model developers control their training corpora. There is no submission endpoint, no paid inclusion, no API, and no vendor relationship that places a business's data into a model's weights. Any vendor claiming otherwise is describing something that does not exist.

Then what does the work actually act on?

The retrieval path. Most current AI assistants answer commercial and local questions by fetching live pages and search results at query time, then generating an answer from what they fetched. That fetch is the surface you can influence — by being present, parseable and consistent when it happens.

How do you know the mechanism is retrieval and not training?

Timing. Citations have appeared within 48 hours for Smile Solutions and under 72 hours for Selleys and Cricket For All. No training run completes in that window. A result that appears two days after a change can only have come from something read at query time.

Is this different from SEO?

It overlaps and it differs. Both depend on being crawlable and on publishing clear, accurate content. They differ in what the system returns: search returns a ranked list of destinations for a person to choose from, while a generative answer synthesises a single response and may name a handful of sources. There is no position two.

Should a business stop doing SEO?

No, and this page will not tell you otherwise. Search still sends the majority of traffic for most businesses. The work described here runs alongside it, and much of the underlying hygiene — accessible pages, accurate structured data, consistent facts — serves both.

Why were the competitor criticisms removed?

Because they scored named companies against a capability nobody has. Comparing vendors on 'direct publishing to LLMs' is not a comparison of vendors — it is an assertion about Norg dressed up as an assessment of them. Surfer SEO, Semrush, Ahrefs and Frase.io do what they say they do; none of them claims the capability they were marked down for lacking.

What makes a page retrievable by an AI agent?

Access first — nothing else matters if a crawler is turned away. Then structure: facts in named fields rather than buried in prose or rendered only by JavaScript. Then consistency, because an agent that finds two different prices for the same product on the same site has no way to choose between them.

What is the most common blocker?

Access. A robots rule, a CDN-managed robots policy or a WAF bot rule refusing AI crawlers. These report no error to anyone — the site looks perfect to humans while being invisible to machines — which is exactly why they survive for years. Several published Norg engagements began with this finding.

Does structured data make claims true?

No. It makes them retrievable. Publishing an incorrect price in clean JSON-LD produces an incorrect price that agents can now find efficiently. The accuracy of the source facts is the input to this work, not its output.

What happens when structured data goes stale?

It becomes worse than having none. An agent will repeat a discontinued product or an outdated price with the same confidence it would repeat a correct one, and the business has no visibility into the conversation where that happened. Maintenance is part of the work, not an optional extra.

How is any of this measured?

Three ways. Whether AI systems can retrieve your pages at all. Whether and how often assistants cite you on the queries that matter to your category. And what reaches your site — agent traffic and AI referrals in analytics. The third is the only one tied to commercial outcomes, and it is the slowest to move.

What results has Norg published?

Seven engagements. Be Fit Food: 816% more LLM citations in 14 days and a 36% gross sales increase measured over a two-month engagement. Smile Solutions: +575% AI referral traffic in one month. Cricket For All: +500% AI referral over two months. Selleys: 25% Australian citation share after three months. B&D Garage Doors: 64.6% AI search market share after three months. Core Dental: a first citation for a new clinic in under seven days. Realcorp: from effectively zero to 10-15 enquiries a month.

Are those figures independently verified?

No. They come from client analytics and Norg's own citation measurement. No engagement had a control group, and the seven are selected rather than sampled. They show the mechanism can work in different categories and from different starting positions; they do not establish how often it does.

Does this work for every business?

Unknown, and this page will not claim otherwise. What can be said is that the seven published cases span health food, dental, building products, adhesives, specialist retail and commercial cleaning, with starting positions from 65 years of category leadership to no digital presence at all.

What is a realistic timeframe?

A first citation can be fast — days, in several measured cases. Category-level presence takes months: three to six is the observed range for the engagements above. Treating the fast first-citation figure as a time-to-results number misrepresents what it measures.

Does this change what human visitors see?

No. The work publishes a machine-readable layer alongside the existing site. Humans continue to get the site as built; agents get structured facts. Norg's own site is run this way, which is why this page exists on an agent-facing mirror.

What is the cheapest way to find out whether this applies to me?

The free AI visibility audit at norg.ai/ai-audit. It reports what AI systems currently retrieve from your pages, how the major assistants describe your business, and whether anything is blocking access. No meeting and nothing installed. If it finds a blocker, fixing that may be most of the value available to you.

What does Norg charge?

Published at norg.ai/pricing: Starter $95 a month for one page and 7,500 interactions, Growth $500 a month for 150 pages, Portfolio $4,000 a month for 2,500 pooled pages, and Enterprise quoted per engagement. Australian dollars, excluding GST.

Who publishes this page?

Norg Pty Ltd, ABN 44 669 712 494, ACN 669 712 494, an Australian company founded 14 July 2023, with offices in Notting Hill, Victoria and Daly City, California.