Once upon a time there was search: ten blue links and a twenty-year contest for first place. Now a growing share of questions no longer produces a list to explore, but an answer to read. Whoever owns the answer owns the market for the question. The thesis of this article is that optimisation for language models — LLMO, Large Language Model Optimization — is not digital marketing’s newest acronym, but the discipline that decides whether an organisation exists in the place where its audiences have moved their questions.
The traffic is not coming back
The numbers describe a transition, not a fashion. The Pew Research Center’s study of real user browsing shows that when an AI-generated summary is present, clicks on a traditional result fall from 15% to 8% of visits, and links inside the summary are opened in 1% of cases (Pew Research Center, 2025). The session, in one case in four, ends on the summary itself.
Nor is the phenomenon confined to consumers: with a quarter of UK businesses already using AI technology — 44% among those with 250 or more employees — according to the Office for National Statistics (ONS, 2026), a growing share of professional questions — suppliers, partners, applications — passes through the same systems.
For those who built visibility on traffic, the signal is uncomfortable: the user served by the answer does not visit, does not browse, does not discover. But the strategic implication is not “defend the clicks”: it is to move the target. The new unit of visibility is not the position in the list — it is the presence in the answer.
SEO optimised for being chosen by a user who sees ten options. LLMO optimises for being chosen by a system that shows one.
How a model chooses
A language model does not “rank” the way a search engine does: it composes. When it answers, it draws on what it learned and on what it retrieves, and it favours — to differing degrees across systems — the sources it can interpret, attribute and reuse without ambiguity.
From this follow the discipline’s operating principles. First: coherence beats emphasis — descriptions aligned everywhere (site, profiles, publications, third parties) make the entity stable in the model’s eyes; contradictory versions make it unreliable. Second: structure beats volume — content organised by questions, with explicit facts, clean definitions and attributed data, is more reusable than ten promotional pages. Third: third-party sources carry weight — being cited by actors the model treats as authoritative counts for more than self-declaration, exactly as in classical reputation.
On this methodological ground the institute has a public foundation: “Fatti trovare da ChatGPT”, the first Italian book devoted to LLMO — published in Italian — written by our founder with AIPIA, the Italian association of AI professionals; the full method is in our publications.
What to measure (and what to ignore)
Serious LLMO begins with measurement, not with content production. A minimum protocol: a basket of questions relevant to your market — commercial, informational, comparative — put periodically to the main systems, in the languages of your markets. Three metrics: presence (are we named?), accuracy (are the facts right?), attribution (are the sources cited ours or third parties’?).
To ignore: promises of a “guaranteed first place in AI answers”. The systems change, the answers vary with the phrasing, and whoever guarantees positions on a non-deterministic mechanism is selling certainties they do not own. The honest discipline works on probabilities: making your facts easier to find, interpret and cite than your competitors’.
For UK organisations there is a circumstance worth stating plainly: English-language answers draw on the deepest and most contested pool of sources anywhere. The reward for vagueness is nil — where your facts are imprecise or scattered, the systems have no shortage of third parties to compose the answer from. Yet in most sectors, few competitors do this work systematically: the advantage goes not to whoever publishes most, but to whoever documents best — consistent, attributable, verifiable facts, maintained over time. Diligence, not budget, is the scarce input.
The line with manipulation should also be stated clearly: the LLMO that works over time is impeccable documentation of true things. Attempts to inflate one’s representation with contrived content have short lives and long tails — the systems get better at recognising them, and reputation does not forgive those who are recognised.
The operating protocol, step by step
For those who want to start this week rather than next quarter, the minimum sequence has five steps.
One: the basket of questions. Twenty to forty questions, built with the people who face the market: the commercial ones (“who can help us with X”), the informational ones (“what is X and who knows about it”), the comparative ones (“X or Y?”), the reputational ones (“whom to trust for X”). The wrong questions are the self-centred ones — “tell me about [our brand]” measures vanity, not visibility.
Two: the survey. The questions must go to the systems your audiences actually use, in the languages of your markets, with the full answers recorded. Answers vary between sessions: several passes over the same question give you the distribution, not the anecdote.
Three: the diagnosis. For each question: present or absent? Accurate or distorted? Cited as a source, or narrated by someone else’s? The summary of the three answers, question by question, is the map of the work to be done — and it usually overturns the priorities marketing would have chosen on its own.
Four: intervention at source. Contradictions first (the conflicting versions of the same information), then absences (the true facts never documented in retrievable form), finally authority (third-party confirmation on the points that matter). In that order: correcting what exists makes what you add more credible.
Five: the cycle. The survey repeated at a fixed cadence, compared against the baseline, priorities revised. The first cycle almost always yields the most visible corrections; from the third onwards the work becomes maintenance — and that is where the discipline pays, because the competitors who did the one-off project are already slipping back.
Who should hold the mandate
Experience suggests a precise division of roles. Execution — surveys, content, corrections — lives well within the communications and digital functions. Ownership does not: the decisions LLMO raises (what to declare publicly, which errors to correct and how, which corporate facts to structure for citability) touch positioning and risk, and require a mandate marketing alone does not have. The format that works: a senior owner, a periodic review in committee, execution where the competence sits.
A subject that precedes marketing
The most common framing error is to treat LLMO as one more channel in the marketing plan. It comes before that: it concerns the factual identity of the organisation — who you are, what you do, what evidence supports it — as the systems reconstruct it. It is the foundation on which reputation in generative engines rests, and for the decisions it touches — what to declare, what to correct, where to invest — it rises quickly to the top.
That is why, in our AI Reputation & Search work, we treat generative visibility as a strategic quantity: it starts with an audit of the representation, priorities are set with leadership, and only at the end is the missing material produced.
What this means for decision-makers
Three moves, in this order. Measure before producing: without a baseline, every investment in content is an act of faith. Correct before promoting: one factual error in the systems’ answers outweighs ten new articles. Assign ownership: presence in the answers is a corporate asset, and assets without owners depreciate.
Search took twenty years to become a corporate function with a budget and named owners. Answers will make the same journey in far less time. The difference, this time, is that latecomers do not lose rankings: they disappear from the conversation.