For twenty-five years the pact was clear: you publish content, Google indexes it, the user clicks, you get the visit. That pact is being rewritten. AI Overviews answer directly at the top of the SERP; ChatGPT, Perplexity and Copilot answer without even passing through a SERP. The user gets the synthesis, and the source receives — when it receives anything — a citation.
The instinctive reaction of many companies is to treat this as a lost-traffic problem. That is a short-sighted reading. The real point is different: source selection has narrowed brutally. A Google page showed ten blue results plus ads; an AI answer cites two, three, sometimes five sources. Whoever makes that shortlist receives a kind of visibility the ten blue links never gave: being recommended, not listed.
For anyone selling products or services, the operational question is one: when someone asks an assistant "what's the best solution for X near me", what does your website need for the answer to contain your name?
AI answer systems, simplifying, do three things: they retrieve relevant documents (often leaning on the search engines' own indexes), they read them, and they compose an answer citing the passages they used. In each of the three steps there is a selection, and in each selection specific traits win.
Clear entities win
Models reason in entities: who is this company, where does it operate, what does it do, what relationships does it have. A website with coherent schema markup (Organization, Service, FAQPage), consistent data across pages and a consistent presence on third-party sources is an entity the model can handle with confidence. A website that describes itself vaguely or inconsistently is noise.
Those who answer win — not those who allude
Models cite passages that answer. A page opening with "Twenty years serving excellence" contains no citable answer. A page opening with "In brief: X is a service that does Y, costs on average Z and makes sense if you are in situation W" is a block ready for extraction. That is why answer-content — definitions, FAQs, summary blocks — gets cited out of all proportion to its weight.
Machine-readable sites win
Clean, fast HTML, sensible semantic structure, an llms.txt file orienting AI crawlers, content accessible without rendering acrobatics. It is not yet a competitive advantage to boast about in a boardroom — it is the equivalent of having the door open when the customer walks by.
Classic SEO optimised for being found. AEO optimises for being taken at your word — by a machine that will repeat it.
In the websites we design, the minimum AEO package consists of five elements, all verifiable from the outside.
First: an "in brief" block on every page that matters. Two or three sentences answering the page's implicit question — what it is, for whom, what it costs or how it works. It is the passage models extract most willingly, and a service to the hurried human reader too.
Second: real FAQs with FAQPage schema. Not the questions the company would like to receive: the ones it actually receives — including the uncomfortable ones, like "how much does it cost" and "how is it different from the competitor". Every FAQ is a potential citable answer.
Third: coherent schema markup across the whole site. Organization with real data, Service for every service, geo-coordinates if the business is local. It is how the company presents itself to machines, and deserves the same care as the brand identity.
Fourth: llms.txt. A text file at the site root telling AI crawlers who you are and which pages describe what. It costs an hour of work. The vast majority of Italian websites do not have one.
Fifth: consistency between what you say and what is said about you. Models cross-reference sources. If the website promises one thing and the reviews tell another, the machine — like the customer — tends to believe the reviews. Citability is also built off-site.
Like every young discipline, AEO already has its snake oil. Three things we see recommended around that produce nothing.
Stuffing pages with "AI". Models do not cite those who declare themselves innovative: they cite those who answer a specific question well. Keyword density died two eras ago; for AI it was never born.
Mass-generating AI content to get cited by AI. A hundred synthetic, interchangeable articles add no information the model does not already have. What gets cited is what is distinctive: proprietary data, first-hand experience, clear positions. A company publishing the numbers of its own observatory has better citation odds than a hundred well-formatted summaries.
Waiting for things to settle. "Let's see how it evolves" is the strategy that missed the SEO train in 2005 and the social train in 2012. Citation patterns are consolidating now, and the sources models learn to trust today will start ahead tomorrow.
The sensible order for an SME is this: first the audit — what ChatGPT, Perplexity and Google answer today when someone asks about your category, and who gets cited in your place. Then the foundations — schema, llms.txt, answer blocks on the commercial pages. Then the distinctive content — the answers only you can give, with your data and your experience. And finally the measurement: AI citations are monitored, exactly as rankings used to be.
None of this requires tearing the website apart. It requires treating the machines that read as an audience in their own right — the audience that, more and more often, mediates with the human one.
Observations based on the AEO/GEO optimisation work carried out on goodea projects and on monitoring ChatGPT, Perplexity and AI Overviews answers for Italian commercial queries, first half of 2026. For an AI visibility audit of your company, write to commerciale@goodea.it.