Generative Engine Optimization: how to be cited, not just ranked
Generative Engine Optimization is the work of making content quotable by AI answer engines — Google’s AI Overviews, ChatGPT, Perplexity, Gemini. It does not replace SEO. It inherits its technical foundations and changes the target: SEO works to have a page ranked in a list; GEO works to have its content repeated inside an answer, with your name attached.
The two do not coincide. You can rank first and never be cited. You can be cited from position eight.
Table of Contents
A note on where this is written from
We are an Italian SEO and communication agency, with offices in Rome and Milan, and our founder has been doing search engine optimisation since 2004 — before the discipline had a settled name in Italy, and two decades before anyone wrote “AI-powered” on a homepage.
That matters here for one reason only: most of what is being sold as GEO today is search discipline with a new label. Teams that arrived with the language models are learning, in public, things that people who worked through Panda, Penguin, mobile-first and the featured snippet already know. Being able to tell the genuinely new part from the rebranded part is the whole job.
So: the genuinely new part is small, and it is real. Here it is.
What actually changed
Where the answer lands has changed. Until recently it landed in a list of links, and the craft was getting into that list as high as possible. On a growing share of searches, it now lands before anyone gets there.
Across the 29 queries we track as representative of our sector, 20 already return an AI Overview (measured July 2026). That is not a forecast. It is the shape of the first screen today. On those queries the user gets a compiled answer, and the links become a footnote to something already read.
The practical consequence: if your content is not built to be extracted, on those queries it does not exist. It is not penalised. It is simply not picked up.
SEO and GEO, side by side
| SEO | GEO | |
|---|---|---|
| Goal | rank a page | get a passage quoted |
| Unit of work | the URL | the paragraph |
| Main signal | links and domain authority | clarity, specificity, verifiability |
| Metric | position, clicks | presence in the answer, brand mention |
| Where it is won | in the SERP | before the SERP |
The second row is the useful one. In SEO you optimise a page. In GEO you optimise one paragraph at a time, because a paragraph is the unit a model lifts. It changes the writing craft far more than it changes the technical craft — which is why so many technically excellent sites are invisible to answer engines.
How an answer engine decides what to quote
Nobody has the formula, and be sceptical of anyone claiming otherwise: there is no direct control over generated answers. There are, however, observable and repeatable behaviours — measurable by querying the engines systematically, which is ordinary empirical work of the kind search professionals have always done.
Extractability. A passage gets lifted if it stands on its own, out of context. Anything beginning “as mentioned above” is unusable to a model.
Verifiable specificity. Between two paragraphs saying the same thing, the one with the number, the date and the source gets quoted. “Many companies” is not quotable. “20 of 29 queries, measured July 2026” is.
Entity coherence. The model must understand who you are, and must find the same answer everywhere: site, schema, profiles, registries, third-party mentions. A company described five ways in five places is a weak entity, and weak entities get named rarely.
Third-party confirmation. What you say about yourself counts for less than what others say about you. This is why GEO reaches into press relations and off-domain citations — it is not confined to your website.
The seven changes that matter
In order of return on effort, based on what we have seen work.
1. Put the answer in the first paragraph. If the summary is not at the top, the rest is not read — by people or by models. It is the least invasive and most profitable change available, because it can be applied to already-ranking content without rewriting it.
2. Write for extraction. Self-contained paragraphs, one idea each, subheadings that are real questions. Every block should survive being copied away from the page.
3. Replace adjectives with numbers. Every generic claim is a missed opportunity. And every number needs a source and a date — an inflated figure gets dismantled in seconds, and at that point ranking is no longer your problem.
4. Mark up structured data. Organization, FAQPage, Article, Author. It does not guarantee citation, but it makes explicit what would otherwise be inferred — and what is inferred is sometimes inferred wrongly.
5. Consolidate the entity. Same description, same details, same references across site, schema, profiles and directories. Tedious, undervalued, and the difference between a company that is recognised and one that is merely present. We know this partly because our own entry took us longer than we care to admit.
6. Update dates honestly. Answer engines prefer recent content. Changing the date without changing the content is a trick that works for one season and costs trust for many: update the content, and let the date follow.
7. Cover questions, not only keywords. People ask models in natural language, in long sentences. Complete questions are worth as much as bare keywords and are far less contested.
How to measure it — and why almost nobody does
This is where the common error hides, and it is not a tooling error. It is a question error.
The first number everyone looks at is how many keywords cite us. It is the wrong number, and wrong in an insidious way: it grows, it looks like progress, and it means almost nothing.
The pattern we meet most often is this. A domain turns out to be cited across a few dozen keywords. You look at which ones, and most of the volume comes from marginal content — something educational or entertaining published years earlier, a side page, a topic with no relationship to what the company sells. The commercial keywords, the ones somebody might buy from, fit on one hand.
The big number is true and useless. The small number is reality.
Hence the KPI we use and recommend: how many commercial keywords the AI cites you on, not how many in total. It is smaller, slower to move, and considerably more honest. Alongside it: the share of your sector’s queries that trigger an AI Overview, brand mentions in answers even without a link, and referral traffic from answer engines — still small, and the only segment growing.
A ten-minute check you can run tonight. Take the list of keywords you are cited on and delete every one you would never sell anything from. What remains is your real starting point.
The mistake we see most often
It is not technical. It is a sequencing mistake.
Almost everyone starts by producing new content. Producing is more satisfying than fixing: a new article is visible, a rewritten paragraph is not. But in most cases the asset already exists: pages ranking around the bottom of page one and the top of page two, on meaningful volumes, uncited only because the answer is buried mid-text. Moving one of those from position eighteen to five, and rewriting its opening so it can be extracted, is worth more than three new articles — at a fraction of the cost.
First make what you have work. Then produce.
Frequently asked questions
Are GEO and SEO the same thing?
No, but they share foundations. A slow, badly structured or unindexable site is cited by no answer engine. GEO adds a layer — extractability, entity, verifiability — that traditional SEO did not require.
Does working on GEO mean giving up SEO?
No. Traditional search still delivers most qualified traffic, and both disciplines rest on the same technical base: a slow or badly structured site neither ranks nor gets cited. The distribution of effort changes; the existence of the work does not.
Can I control what AI says about my company?
You cannot control it. You can influence it, by making the information available about the company coherent and verifiable. Anyone promising control over generated answers is promising something that does not technically exist.
How long does it take to see results from GEO work?
On already-ranking pages, the first movements appear within weeks. On new content, timescales remain those of SEO: a few months. Which is why content that must be visible in September is published in August.
Do I need an agency to work on GEO?
For some of it, yes; for other parts, no. Marking up structured data, rewriting paragraphs so they can be extracted and updating existing pages are things an in-house team can do, and they already deliver part of the result. Entity consolidation — making the information about a company consistent everywhere it appears — and systematic measurement of citations need tools and continuity: that is usually where internal effort stops.
We have been optimising for search engines since 2004, and for answer engines since answer engines existed. We did not discover this field when it became fashionable, which is the only reason we can tell you which parts of it are genuinely new.
If you want to know how many commercial keywords currently cite your company — and how many could — a GEO audit is where it starts.
