For twenty years digital reputation had an address: the first page of search results. Organisations learned to guard it, measure it, defend it. That address is emptying. When a prospective client, an investor or a journalist asks an AI system “who can I trust for X”, the answer is not a list of links: it is a summary — with names included and names omitted. The thesis of this article is that reputation in generative engines is already a real quantity, forming with or without your contribution, and that guarding it has become a matter for senior leadership.
The evidence of the overtaking
The data on user behaviour are stark. According to the Pew Research Center’s study of the real browsing of a sample of US adults, when an AI-generated summary appears on the search page users click a traditional result in 8% of visits, against 15% when there is no summary; links cited inside the summary are opened in 1% of cases (Pew Research Center, 2025). A quarter of sessions end directly on the page carrying the summary: the user has the answer, the journey stops there.
The consequences are two, and the second runs deeper than the first. The first: less traffic to your content. The second: the representation matters more than the destination. If the system summarises you badly — or does not name you at all — the user will never reach the site where the story is told well.
In traditional search the risk was not being found. In generative engines the risk is being told badly — by a voice the user considers neutral.
How the representation forms
A generative system builds what it says about an organisation from three layers: what it learned in training, what it retrieves at the moment of the question, and how the two combine in the summary. None of the three is directly controllable. All three can be influenced.
The retrieved layer is the most sensitive: coherent, verifiable, well-structured sources are more likely to be cited. Inconsistencies — the biography that says one thing, the profile that says another, the press release that contradicts both — do not produce an average: they produce an unstable account that changes with every question. And omissions weigh as much as errors: where your facts are absent, the summary is built from other people’s.
This guarding work has a method which, in its operational detail — how language models select and cite sources, and how to structure your own so they are selectable — is the terrain of LLMO, the discipline born for exactly this.
A scenario to grasp the stakes
Picture a solid British manufacturer, ninety years of history, never a material dispute. An international fund evaluates it for an acquisition, and the analyst — as everyone now does — opens the file by asking a generative system for a profile of the company. The summary is accurate on almost everything — but it links the firm to a product recall from twelve years ago, told without its ending (the recall was voluntary and unfounded), because the sources the system retrieves are three articles from the time and no later correction.
Nobody in the company ever knew of that representation, because nobody had ever put that question to that system. The analyst did. The cost is unmeasurable — a negotiation that opens under a shadow, an extra question in due diligence, an implicit discount on the price.
The scenario is constructed, but every element of it is ordinary: dated information prevails where current information is missing, omissions weigh as much as errors, and decision-makers use these tools far more than they declare — organisational AI adoption has reached 88% of the entities surveyed by Stanford University’s AI Index (Stanford HAI, 2026), and background files on suppliers and counterparties are among the first uses to become routine. Generative reputation is not a subject for the future: it is a file already being prepared, without your knowledge, about you.
The three levels of guarding
Level one: the baseline. The inventory of answers described below, repeated at a fixed cadence — quarterly for most organisations, monthly for those in exposed sectors. Without a baseline, every discussion of generative reputation is anecdote.
Level two: source coherence. The systematic review of what the systems can retrieve: website, institutional profiles, public registers, publications, press coverage. The aim is not to produce new content but to remove the contradictions among what exists — the least glamorous and most profitable work in the whole discipline.
Level three: the response to deviations. A protocol for when measurement finds a problem: who assesses severity, who corrects at source, who checks that the correction propagates. Corrections in generative systems are neither instant nor guaranteed — one more reason to treat them as a process, not an emergency.
From a marketing matter to a governance matter
As long as the conversation is about commercial visibility, the subject can stay within marketing’s perimeter. But generative representation touches ground marketing does not govern: how the systems describe your legal position, your solidity, your conduct. An inaccuracy repeated by a voice perceived as neutral is not a traffic problem: it is a reputational risk with potential legal and financial tails.
That is why the question “what do AI systems say about us?” deserves the same treatment as “what does the press say about us?”: continuous guarding, periodic measurement on the questions that matter, an owner with a mandate. It is the perimeter of the AI Reputation & Search work the institute carries out with organisations — analysis of how the systems represent the brand, and strategic direction of its presence.
Measure before reacting
Guarding begins with an honest inventory. Choose the twenty questions a client, a candidate or an investor would ask about you and your market. Put them to the main systems, in several languages if you operate in several markets. Record: are we named? Are the facts correct? Are the sources cited ours or third parties’? What is the tone of the summary?
For organisations operating across markets, the measurement must be done language by language: the systems do not answer the same way in English, French and German, because they retrieve different sources and weigh different authorities. A British organisation impeccable in its English-language answers may be invisible — or worse, told only by third parties — in the languages of its export markets and investors. Generative reputation is not one: it is one for every language in which someone looks for you.
The result is almost always instructive: dated representations, competitors cited on questions where you should appear, factual errors nobody had ever gone looking for. From there the work becomes concrete — correcting inconsistencies at source, structuring the content that is missing, making citable what is true. It is not a one-off project: the representation drifts, and it must be watched as one watches the press cuttings.
What this means for decision-makers
Three decisions to take now. First: assign ownership — generative reputation cannot belong to nobody, because in practice it already belongs to someone else: the systems. Second: put the measurement into the reporting cycle, alongside brand awareness and press review. Third: treat wrong answers as one treats errors in print — with documented correction at source, not with indignation.
Reputation has always been what others say about you when you are not in the room. Today the room contains a generative system, and it speaks to millions of people a day. It is worth knowing what it says.