AI 101 — What is generative engine optimization (GEO)?

Generative engine optimization is the practice of shaping a web page so that AI answer engines — ChatGPT, Google's AI Overviews, Gemini, Claude — quote it and link to it when someone asks a question. It is search engine optimization for a search box that writes an answer instead of handing you ten links.

You have probably met it without the name. Someone asks ChatGPT a question about your industry and the reply cites four websites, three of them competitors. GEO is the set of habits people adopted in response — and, as the research now shows, many of those habits are close to useless.

Why it matters right now

The size of the shift is measurable. A study of Google's AI Overviews, published on arXiv in February 2026 by Mehrzad Khosravi and Hema Yoganarasimhan, compared English Wikipedia's search referrals against the same articles in German and French as AI Overviews rolled out country by country. It found default availability cut English search traffic by 5.45% and 4.82% respectively, per the paper. Fewer people click through, because the answer is already on the page.

The citation side is now studied just as closely. A September 2026 analysis tracked roughly 2 million AI citations across four commercial engines — ChatGPT, Claude, Google AI and Gemini — over six months, tied to 10,000 crawled pages from 19 business-software companies. Its headline finding: the page feature that best predicts being cited is how closely the page's wording matches the queries a site's audience actually asks, according to the study's authors.

That matters to anyone who owns a website, not just marketing departments: the traffic these engines send is smaller, and it arrives from a summarizer that decided your page was the best material for a sentence it was already writing.

The mental model: three gates, not one ranking

Classic search is one gate. You rank, a person clicks, they arrive. An answer engine runs a longer assembly line, and a page can fail at any stage.

First comes retrieval: does the engine's search step find your page at all? Second comes use: is your sentence worth quoting, or does it lose to a competitor's clearer one? Third comes the citation: does the engine name you, link you, or silently absorb the fact? A page can be retrieved and quoted with no link shown — the case the Wikipedia study measures, and where most of the lost traffic lives.

The engines are also becoming more search-like rather than less. Simon Willison documented ChatGPT's search using site-restricted queries at scale, writing in August 2026, which means the retrieval half of the pipeline behaves much like an ordinary index — and the ordinary blocking rules, robots files and access controls, still decide who gets in.

The analogy: the librarian who writes the summary

Picture a library where you never walk the shelves. You ask the librarian a question, she disappears into the stacks, reads six books, and hands you one paragraph plus three titles she thought were worth naming.

You cannot charm your way onto her list with a nicer cover, or with a table of contents that promises everything. Two things work. Your book has to answer the questions people bring to that desk — and it has to be one she trusts, from a shelf she has learned to rely on. Everything else is decoration.

What the evidence says actually works

The first GEO paper, from a Princeton-led team and published at KDD 2024, formalized the field and reported that the right content edits could lift visibility in generative answers by up to 40%, the authors found. The detail is duller and more useful than the headline: the winning moves were citing sources, adding statistics and quoting experts — things that make a paragraph checkable, not things that make it long.

The 2 million-citation study sharpens this considerably, and partly contradicts the industry's advice. It tested more than 60 page features and found that the standard answer-engine checklist — FAQ blocks, structured data, fast page loads — looked positive in pooled data but shrank to nothing, or reversed, once the researchers controlled for which domain the page lived on. They describe it as a Simpson's paradox with practical consequences. Domain-level AI authority, meanwhile, outweighed the strongest page-level feature by a factor of six.

Translation: the engine is mostly deciding whose site to trust before it reads the page. A thin page on a site models already cite will outperform a beautifully optimized one on a site nobody has heard of.

Common misconceptions

"It is just SEO with a new name." Related, not identical. Keywords still matter, but the unit of competition is the paragraph that gets lifted into an answer, and the winning trait is verifiability — numbers, named sources, specific claims — not density.

"There is a trick that forces citations." The tricks tested — schema markup, FAQ spam, listicles rewritten into question form — mostly wash out in careful studies. What survives is being the page that already answers the question in plain sentences.

"Being cited means being paid." Not necessarily. The Wikipedia result is the clearest warning yet: citations rose while clicks fell. A citation is branding, not traffic, and teams that report only citation share are measuring half the outcome.

"AI answers are a fad and the traffic comes back." The referral numbers have moved in one direction for two years. Plan for less click-through and better attribution, not for a return to 2019.

Where to learn more

Read the two primary papers above — they are unusually readable, and both publish their methods. If you own a site, the practical starting point is smaller than any vendor will tell you: find the twenty questions your customers ask, and make sure one page each answers them directly, with a number and a named source in the first two paragraphs.

For the retrieval half of the assembly line, our explainer on what RAG is covers how a model is handed a set of documents to read before it answers. And if you are wondering whether volume still helps, what AI slop is is the reason it no longer does — engines increasingly filter the pages that look machine-made at scale.

Should publishers be paid when an AI answer quotes them? Write to us via our contact page.

AI 101 — What is a vector database?

A vector database is a database purpose-built to store and search "embeddings" — long lists of numbers that capture the meaning of a piece of text, an image, or audio — so it can find things that are semantically similar rather than things that merely match your keywords. If that sounds abstract, consider where you have already met one.

AI 101 — What is tokens per second (tok/s)?

Tokens per second is how fast a language model writes: the number of word-fragments it produces each second once it has started answering. It is the spec that has quietly moved to the top of every model release. Context windows used to be the number vendors fought over; now the second line of the announcement is a speed claim, usually with a hardware name attached and a batch size buried in the fine print.

AI 101 — What is a KV cache?

A KV cache is the working memory a language model keeps of the text it has already read, so it can write the next word without redoing the work for every word that came before. That is the short answer. The reason it deserves an explainer is that this cache — not the model's weights — is usually what decides how many people one GPU can serve at once, and therefore what long conversations and agent loops actually cost. Why it matters right now Every agent that runs for twenty minutes and…

AI 101 — What is a GPU, and why does AI need one?

A GPU is a computer chip built to do thousands of small arithmetic operations at the same time — which turns out to be the exact shape of the math inside a modern AI model. The name stands for graphics processing unit, because the chip was designed to draw video games frames. AI is the job it turned out to be extraordinary at. If you follow AI news, the GPU is the noun that never leaves the sentence: who is buying them, who cannot get enough of them, why one chip costs more than a car, why…

AI 101 — What is context engineering?

Context engineering is the work of deciding what text a model sees before it answers: the instructions, the retrieved documents, the tool descriptions, the conversation so far, the notes from last week — and, just as importantly, what to leave out. Prompt engineering is one slice of it. Context engineering is the whole tray. Every time a model writes a reply, it sees exactly one flat block of text and nothing else.

AI 101 — What is an AI browser?

An AI browser is a web browser with an AI model wired into the browser itself, not bolted on as a separate chat tab. The model can see the page in front of you, the tabs you have open, and sometimes your history and logins — and, if you let it, it can act: click, type, fill forms, compare prices, book things. That definition hides the split that actually matters. Reading is one capability. Acting is another.

AI 101 — What is a chiplet?

A chiplet is a piece of a processor that is manufactured as its own small chip, then packaged alongside other small chips so that the finished part behaves like one big chip. Instead of printing one enormous processor, you print several modest ones and wire them together inside a single package. Every flagship AI accelerator shipping today is built this way. It matters right now because the constraint on AI hardware has moved.

SolarWinds puts IT agents on a deterministic leash

Enterprise IT is where AI agents stop being a demo — a wrong guess there costs uptime, audit findings, and weekend callouts. SolarWinds' answer is to let the model decide what needs doing and never let it decide how far it can go. SolarWinds today launched Avior, a standalone agentic AI system that takes on first-line IT operations work inside a fixed, auditable execution framework — the company's answer to the reliability problem of letting probabilistic models act on production systems.