.png)
Key takeaways
- Generative engine optimisation was named in a 2023 research paper, which found that citations, statistics, quotations and authoritative language improve AI visibility by up to 40%. Keyword stuffing did not work at all.
- The mechanism differs from search: SEO logic is rank higher and get traffic, while generative systems extract and cite, so a lower-ranked page with better evidence can win.
- Google says there is no such thing as GEO separate from SEO. Bing uses the term. Forrester says the difference is significant but not fundamental. The disagreement is real and unresolved.
- Everything the research says works is something an editor would have demanded before the internet existed: name your sources, show your numbers, quote the people who know.
- Whether that pattern holds outside search is an open question, but it is worth asking.
What the research actually found
The term "Generative engine optimisation" (GEO) was named in a November 2023 paper from researchers at Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi. It demonstrated that targeted optimisation could increase visibility in generative responses by up to 40%.
The interesting part is not the number. It is the list of what worked.
Citations. Statistics. Quotations. Authoritative language. The study also tested keyword stuffing, which did not improve results at all.
That list would be unremarkable in a newsroom. Name your sources. Show your numbers. Quote the people who know more than you. An editor would have sent a piece back for failing any one of them, and would have been baffled to hear it described as an optimisation technique.
Why the mechanism is different
There is a real structural change underneath this, and it is worth being precise about it.
Traditional search logic runs: rank higher, get traffic. Generative systems do not rank pages, they extract and synthesise information from multiple sources into a single answer. So the logic becomes: provide better evidence, get cited.
The practical consequence is that a lower-ranked page can dominate an AI answer if its content is more structured and evidence-rich than the pages above it. Position matters less. Being quotable matters more.
That is a genuine shift, and anyone telling you nothing has changed is not paying attention.
The argument nobody has settled
Whether this constitutes a new discipline is genuinely contested, and the disagreement runs through serious people.
Google released official documentation in 2026 stating there is no such thing as generative engine optimisation separate from SEO, the first time it had committed that position to writing, having made the point at conferences beforehand. Bing, by contrast, uses the term in its own documentation while noting that SEO fundamentals support eligibility for AI-generated experiences.
Forrester's Nikhil Lai argued that answer engine optimisation and its cousins are significantly but not fundamentally different from SEO, and that advocates of the various acronyms tend to exaggerate the differences to carve a startup-sized hole in marketers' tech stacks. Venture capital has flowed into the category regardless.
Pulling the other way, the studies suggest search rankings and AI citations really have come apart. Brandlight's analysis, synthesised across 680 million citations from five major AI platforms, found the overlap between Google's top-ranking pages and AI-cited sources fell from roughly 70% in early 2024 to under 20% by April 2026. Ahrefs put the figure at 12%.
Moz, across 40,000 keywords, found 88% of AI Mode citations did not come from the organic top 10. Andreessen Horowitz partners have declared the end of SEO outright.
Though it is worth noting those numbers do not agree with each other, and one that cuts the other way: Rankability found that pages ranking in positions one to three were cited 98.9% of the time, against 49.3% for pages outside the top ten. Ranking well still helps. It just no longer guarantees anything.
The most useful finding I came across was neither of those. An Ahrefs study of 75,000 brands found that branded web mentions correlate with AI visibility at 0.664, while backlinks manage 0.218. Muck Rack, analysing over a million AI-cited links, found 82% came from earned media.
Which is to say: the strongest signal is being talked about, credibly, by people who are not you. That is not a link-building tactic. It is public relations, and before that it was just having a reputation.
The search consultant Will Scott lands closest to where I've ended up, and more concisely than I managed.
"What I’ve been saying is, yeah, [GEO is] just SEO if you’ve been doing good SEO. But the problem is so much SEO that gets done is still the same old thing."
- Will Scott, CEO & Co-founder at Search Influence
That is not a dismissal. It is the whole point.
E-E-A-T, and what it is not
Google's framework of Experience, Expertise, Authoritativeness and Trustworthiness (E-E-A-T) gets treated as a dial you can turn. It isn't one.
E-E-A-T is defined in Google's Search Quality Rater Guidelines, the handbook used by human reviewers assessing search results. It is not a confirmed direct ranking factor. It launched as E-A-T in 2014, and the second E (Experience) was added in December 2022 to prioritise content rooted in first-hand knowledge.
Trust sits at the centre of the four. The guidelines are explicit that an untrustworthy page has low E-E-A-T no matter how experienced, expert or authoritative it appears, and they instruct raters that when a website's claims about itself conflict with what independent sources say, the independent sources win.
That addition is the tell. Google looked at the web, saw an ocean of competent, sourceless, experience-free content, and added a criterion that amounts to: has the person writing this actually done the thing?
Which is, again, a question every editor has asked since editors existed.
What this might mean elsewhere
Here I am speculating rather than arguing from evidence, and I want to be clear about which is which.
The pattern in search looks like this: an industry optimised away from craft for twenty years because the machines rewarded something else, and then the machines got better and started rewarding craft again. The correction did not come from writers winning an argument. It came from the technology changing its mind.
I suspect that pattern is not unique to search. Plenty of industries have spent the digital era optimising for measurable proxies rather than for the thing the proxy was standing in for. If the tools are now good enough to assess the real thing, some of those optimisations become liabilities rather than advantages.
But I would not want to push that further than it goes. One paper about citation behaviour in language models is not a theory of everything, and "we should go back to how things were" is the kind of claim that feels true more often than it is true.
What I would actually do
If you are writing for a technical company and want to be cited by AI systems, the advice is dull and it is the same advice a subeditor would have given you:
Cite your sources, and link to them. Use real numbers rather than vague claims about growth. Quote people who know the subject, by name. Put an author on the page with a bio that says why they are worth reading. Structure the piece so a specific claim can be lifted out of it without losing its meaning.
None of that is new. That is rather the point.
About the author
Kyle Smith has spent a decade explaining blockchain infrastructure, cybersecurity frameworks and AI adoption to people who don't build them. He is Marketing Manager at Miraculous Equation, founded the digital agency Essential Millennial, and started out in newsrooms. Based in Braga, Portugal.
.png)
