28 August 2026 · 8 min read
Almost every argument about generative engine optimisation comes from conflating two completely different things: what a model learned when it was built, and what it fetches when it answers.
Short answer
Generative engine optimisation improves how likely an AI system is to use and cite your content. It works through two distinct mechanisms: training, which is fixed until the next model and cannot be influenced quickly, and retrieval, which happens live when an assistant searches the web and is genuinely influenceable today.
01 The mechanics
Training is what a model absorbed when it was built. If your business was not well described on the web at that point, nothing you publish today changes it — the corpus is fixed until the next model ships. This is why new businesses are invisible to assistants and why patience is not optional.
Retrieval is different. When an assistant searches the web to answer a question, it fetches pages now and quotes them now. That is where work you do this month can land, and it is why crawler access and current, clearly written pages matter more than most GEO advice suggests.
The term itself comes from a 2023 academic paper by researchers at Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi, which measured which changes to a source page altered its visibility inside generated answers. It found additions such as quotations, statistics and cited sources helped, while conventional keyword optimisation did comparatively little. That is evidence about specific systems at a specific time — considerably more than most GEO advice rests on, and not a permanent rulebook.
02 Implications
Each of these is a consequence of how retrieval works, not a guess about a ranking factor.
03 Scepticism
Apply these to anything sold to you as generative engine optimisation, including by me.
— Questions
It shares a foundation — crawlable, clear, credible content helps everywhere. What genuinely differs is the emphasis on entity consistency across sites you do not control and on third-party corroboration rather than ranking. It is closer to public relations than classic SEO.
Not on demand. Training happens at intervals on a corpus you do not control. What you can do is be well described and widely corroborated now, so the next training pass has good information about you.
That is what the 2023 research found on the systems it tested. It is consistent with how retrieval works, since specific attributable content is easier to quote safely. Treat it as good evidence rather than a permanent law.
On the durable parts, yes — consistent descriptions, correct schema, a deliberate crawler policy, one genuinely citable page. Those help conventional search regardless. A large dedicated GEO budget for a small business is, in my view, premature.
— More
AEO & GEO · 6 min
This is the single most common technical misunderstanding in AI search right now, and acting on it incorrectly can remove a business from Google Search altogether.
ReadSEO · 7 min
It is still one of the most commonly sold local SEO tactics in India, it works briefly, and it is a named spam pattern in Google’s own documentation.
ReadGoogle Ads · 6 min
Almost every underperforming account I open has the same three problems. None of them are exotic, all of them are visible within an hour, and fixing them usually matters more than any bid strategy.
ReadFree consultation · No obligation
The reasoning on this page is public because I would rather be judged on it than on a claim. If you want it applied, that is the service.
Name, number, email, what you need.