How AI builds a picture of your organisation, and what to do about it.
Listen to this article · 21 minutes
Read by an AI voice.
Before a customer, investor or candidate visits your website or speaks to anyone who works for you, they can ask an AI tool what you are like. The answer draws on years of public material, including pages your communications team has never seen. It may be accurate or out of date, kind or harsh. Either way, it may be the first account of your organisation that person reads.
AI reputation is the picture of an organisation that people get when they ask an AI tool about it. Some call it reputation in AI. The picture comes from what the AI has absorbed or can look up: what the organisation says about itself, what it visibly does and what other people say about it. It can be accurate, incomplete, out of date or wrong, and it can change from one tool or one question to the next.
It matters because people use it to decide whether to buy from, work for, invest in or trust an organisation.
It also changes the job. Reputation work has long meant controlling the message. An AI tool reads far more than your message. It compares what you say with what you do and with what others write about you. The response is signal alignment: what an organisation says, what it does and what others say about it should point the same way. The rest of this article explains how AI builds its picture of your reputation and where your influence over it stops.
More people now turn to AI when they research an organisation before making a decision.
Gartner surveyed 645 business buyers in August and September 2025. Forty-five per cent had used generative AI in a recent purchase, mostly to research suppliers and products. Sixty-nine per cent said they preferred to check what AI told them with a salesperson. AI is entering the research stage while people stay involved in the decision.
Journalists are using it too. Muck Rack's 2026 State of Journalism survey, which drew mostly on US journalists, found that 82 per cent use at least one AI tool, up from 77 per cent a year earlier. Muck Rack sells software to PR teams, so treat the figure as indicative. It measures AI use in general, not research on companies.
The public is moving the same way. The Reuters Institute's 2025 report found that weekly use of AI tools such as ChatGPT rose from 18 to 34 per cent between 2024 and 2025 across six countries. Of people who had seen an AI-generated search answer in the past week, only about a third said they always or often click through to the sources behind it.
The research does not yet show how often investors and job candidates do the same, but nothing stops them.
For communications leaders, this means an organisation's first impression may come from an answer it did not write, built from information it does not fully control.
An organisation can show up in an AI answer and still be misrepresented.
Much of the talk about generative engine optimisation (GEO) is about whether companies get mentioned or cited in AI answers. That is a fair aim. It leaves out a second question: what does the answer say about the organisation?
Take a hypothetical software company. It spent its first years selling to small businesses, then moved to large enterprise customers. Its website now describes the new position. But older articles still describe the small-business product, and an industry directory lists services it dropped.
Ask an AI tool about it and the answer may cite three credible sources and describe the company as it was two years ago. The company is visible. The impression is wrong.
The same thing happens when an answer names a company correctly but picks a one-sided selection of its work, customers or controversies. A citation shows where the AI looked. It does not show that its conclusion is complete or right. An accurate source can still support an out-of-date story when nobody notices its date.
AI tools answer questions about organisations from two places. The mix depends on the tool and the question.
The first is what the AI absorbed in training. It has read huge amounts of text, including material about companies, executives and events, and it holds on to patterns from it. So it may already associate your organisation with certain people or stories before it looks at anything current. What it knows can be incomplete or out of date. A recent acquisition, leadership change or corrected allegation may be missing.
The second is what it looks up. Some tools search the web while they prepare an answer. They find pages that look relevant, use some of them and write a response. ChatGPT with search switched on is one example, and OpenAI itself says its search results and citations can be incomplete, outdated or incorrect.
Either way, the answer arrives as one smooth explanation. Smooth is not the same as complete. The AI may have read only a few pages, missed evidence that points the other way or leaned on whatever was easiest to find. And you cannot see how much weight it gave each source.
So when you read an answer about your organisation, ask what the AI had to work with, which evidence shows up in the answer and how closely it matches the real record.
Your reputation in AI can be read through three layers of evidence: what the organisation says about itself, what it appears to do and what others say about it. This is Adelbury's framework, and it is the basis of Adelbury's AI reputation audit. It is a framework for reading an organisation's public footprint. It does not claim that AI tools sort their sources into these three groups.
Declared narrative. This is what the organisation says about itself: its website, leaders' biographies, company profiles and the way it describes what it stands for. It shows how the organisation wants to be seen. It does not prove any of it is true. The software company says it serves large international customers and stresses security and stability.
Observed behaviour. This is what the organisation appears to be doing in practice: the jobs it advertises, its product pages and pricing, what its leaders say in public and what its people post. This kind of evidence is harder to polish, but it still needs careful reading. A job advert does not prove someone was hired, and a product page does not prove the product is good. The software company's adverts are mostly for people to build basic enterprise features, and its product pages describe features that are still being built.
External validation. This is what other people say: independent journalism, reviews, public filings, academic research, trade bodies, specialist publications and online communities. These sources can back up, qualify or challenge the organisation's own account. They are not automatically right. Reviews can be unrepresentative, news stories can be old and directories contain mistakes. Customer reviews of the software company repeatedly mention problems with complex set-ups.
The most useful findings come from comparing the layers. None of this proves the company's enterprise claim is false. It may be growing into the position or fixing known weaknesses. But taken together, the evidence raises questions the website does not answer. An AI tool drawing on this material could paint a less favourable picture than the website does. It could also be unfair if the evidence were old or lopsided.
A proper review would first establish what the record shows. Only then can you tell whether the problem is weak communication, a real gap in what the company delivers or a mistake in how the AI read the evidence.
That is the argument at the centre of Adelbury's framework. An organisation's reputation is more credible when what it says is backed by what it visibly does and by what others independently say. Where the layers disagree, investigate. Closing those gaps is signal alignment.
How much a source counts depends on the question, the AI tool and what was available at the time.
Adelbury's AI reputation audit looks at sixteen types of source, grouped by layer. Each has an indicative influence rating. The ratings are Adelbury's working judgements. They are not measured weights and not a universal ranking.
| Layer | Source type | Indicative influence |
|---|---|---|
| Declared narrative | Corporate website | Medium |
| Declared narrative | LinkedIn company page | Low to medium |
| Declared narrative | Leadership biographies | Medium |
| Observed behaviour | LinkedIn activity | Medium to high |
| Observed behaviour | Job listings | High |
| Observed behaviour | Executive commentary | High |
| Observed behaviour | Product and pricing pages | Medium |
| Observed behaviour | Employee activity | Medium |
| External validation | News and editorial coverage | High |
| External validation | Podcasts and interviews | High |
| External validation | Independent and creator media | Medium to high |
| External validation | Reviews | High |
| External validation | Forums and communities | Medium to high |
| External validation | Wikipedia | High |
| External validation | Directories and filings | Medium |
| External validation | Institutional and authority sources | High |
A few of these deserve a note. Job adverts show where a company plans to invest and which skills it wants. They show intent, not proof that the capability exists. What leaders say in public attaches claims to named people, which is why it often counts for more than a corporate statement. Independent journalism can confirm or challenge what an organisation says, and trade press may matter more than national coverage on a sector question. Reviews record what people experienced, and their value depends on whether they are credible, representative and recent. Company filings, regulator publications and academic research can establish facts that promotional material cannot.
Sources also do different jobs. A company filing may settle who owns a business and say nothing about whether customers are happy. An interview may explain a leader's strategy but cannot show it is working. So the source that matters most depends on the question. A question about ownership may draw on filings, one about workplace culture on reviews and a comparison of expertise on interviews, published work and qualifications.
Nothing shows that every AI tool looks up or favours the same sources. The table is most useful for spotting gaps in the public evidence about an organisation and deciding which records to look at first.
A wrong or unfair AI answer can have four different causes. Each needs a different response.
The AI made something up. AI tools sometimes state things that are false but sound convincing, including invented facts and sources. OpenAI's own research says this happens partly because the way models are trained and scored rewards a confident guess over an honest "I don't know". PAN Communications reported in February 2026 that of more than 11,000 links ChatGPT gave in response to executive-level research questions, 69 per cent were real and correctly credited. The rest pointed to the wrong source (19 per cent) or to sources that do not exist (12 per cent). It is one commercial study, so treat it as indicative. When an AI invents a claim there may be no page to correct. Record the error, check whether it keeps happening and report it to the platform.
The evidence is out of date. An old biography or product description may have been right when it was published. In the software company's case, older articles still describe a small-business product. Updating the sources you can update helps, though it does not guarantee that every AI tool will pick up the new version.
The evidence is thin or hard to find. An answer can be wrong because there is little relevant material, or it is hard to reach, not because anyone published something false. Making current, authoritative information easier to find gives the AI more to work with.
The record contradicts itself. What a company says may differ from what its products, decisions or customers show. If reviews keep reporting problems with complex set-ups, editing the website will not settle it.
Do not mix these up. An invented allegation needs a different response from a criticism that has a source. And when an organisation's preferred story disagrees with an accurate record, the AI has not necessarily made a mistake.
AI answers are not fixed. They change with the tool, the way the question is worded, the sources available and the conditions at the time.
Different tools use different models and search services, and updates change how they behave. A general question about a company may return a summary of its products and leaders. Ask about risks or how it treats staff and a different set of evidence comes into play. Place matters too. OpenAI says ChatGPT may use an approximate location, based on your IP address, to provide relevant local results.
Research on brand recommendations shows how far this can go. In late 2025, SparkToro and Gumshoe.ai ran 12 recommendation questions 60 to 100 times each on ChatGPT, Claude and Google's AI features. The same list of brands came back in fewer than 1 per cent of runs. The study was run with an AI tracking company, has not been peer reviewed and looked at recommendation lists rather than answers about reputation. It still shows why one answer proves very little.
For an organisation, that variation is something to measure. One good answer tells you little about what others will see. Only repeated questions, across several tools, wordings and dates, show which descriptions keep coming up and which depend on the circumstances of one search.
You can improve the evidence AI tools have to work with. You cannot dictate the answers.
What you can do depends on what is going wrong, and a few searches will not tell you that. Once you know the cause, the options are clear. Where a source is wrong and there is a proper way to ask for a correction, use it. Where your own information is incomplete, publish clearer facts and the evidence behind them. Where an AI invents something, use the platform's error reporting and keep examples so you can compare later. Where the record contradicts what you say, the fix lies in how the business runs, not in the wording.
These steps make the public record more accurate and easier to find. They do not guarantee that a particular source will be picked up, cited or believed. And no company can expect to remove every unfavourable account. Independent reporting, fair criticism and historical facts stay in the record.
This approach also needs a public record to work with. It works best for organisations with a substantial public footprint. If the record is thin or new, the first job is to build credible evidence, because the AI has little to compare.
The response needs editorial judgement as well as technical work. Some findings call for a correction. Others need more evidence, a fuller explanation or a change in how the organisation behaves.
Communications is well placed to coordinate AI reputation, because the evidence behind it is spread across the organisation.
Marketing runs the website. HR owns recruitment pages and staff communications. Product and commercial teams own the product pages and help content. Investor relations and legal handle financial disclosures. Executives speak for the organisation in public. Each team produces material that feeds the outside picture, often without thinking about how it fits with the rest.
Communications already looks at how an organisation's actions and statements shape what people think of it. AI widens the material to cover and adds another route by which people meet it. The job is to coordinate, investigate and use judgement. It might mean making sure an acquisition appears consistently across company descriptions and public records, reconciling two accounts of one product or spotting repeated customer complaints that contradict a published promise.
There is a limit. If an organisation claims excellent customer service while current, credible evidence shows persistent failures, the problem lies in how the business runs. A better-written account will not settle it.
This is why the shift from message control to signal alignment matters. Communications cannot control every signal. It can see across them, bring together the people who own the facts and flag where the record disagrees with the story.
Start with a proper audit, not a handful of searches. A few searches show what one AI tool said on one day. They cannot show whether the answer is stable, where it came from or what to do about it.
A professional audit has four parts.
The tools collect the answers. Judgement decides what they mean: which conclusions could affect trust or a commercial decision, which evidence is producing them and whether the fix belongs with communications, leadership, HR, product, legal or investor relations.
That is what Adelbury's AI Reputation Audit does. It combines professional tools with Adelbury's framework to give you a tested view of how the main AI tools describe your organisation, a map of the sources behind it, an assessment of where your three layers agree and disagree, a ranked list of risks and opportunities and an action plan with owners. The report behind it, The AI Reputation Audit, sets out Adelbury's framework and the sixteen-source matrix in full. To arrange an audit for your organisation, see what it includes.
What is an AI reputation audit? explains the process in more detail.
No. Generative engine optimisation (GEO) is mostly about getting an organisation mentioned in AI answers. AI reputation asks what those answers say and whether the picture is accurate, credible and backed by evidence. The two overlap, but being mentioned does not mean being well regarded.
They are closely related. Online reputation management deals with how an organisation appears across websites, search results, news, reviews and social media. AI reputation adds a further question: how AI tools combine and interpret all of that when they answer.
No. No organisation can guarantee what an independent AI tool will say. You can improve your public information, correct errors where possible and publish verifiable evidence. Those steps may influence future answers. They do not give you control.
It can. PR produces independent coverage, public comment from leaders, corrections and relationships with credible outside sources. All of these form part of the record AI draws on. It also gives communications teams a reason to check that public claims match what the organisation does and what others can verify.
It depends on the organisation's profile, the decisions people make about it and how fast its record changes. Monthly or quarterly checks are a reasonable start, with extra checks after major announcements, leadership changes or controversies. Compare results consistently rather than relying on a single answer.
Stephen is the founder of Adelbury. He has spent the last twenty years working in strategic digital communications for some of the largest agencies and brands.
