AI Reputation Management: What It Is and Why It Is Now a Comms Job

Somebody is deciding whether to buy from you right now, and they are not on your website. They are asking an assistant whether you are any good. They will get a confident paragraph back, written in a friendly voice, citing nothing or citing three sources you have never audited. They will believe most of it. And nobody on your team will ever know the conversation happened.

That is the problem AI reputation management exists to solve. It is the work of finding out what AI assistants say about your brand and changing it. You cannot log into a model and edit the paragraph, so the job is done one step back, on the sources those models read and repeat. That single fact is what makes this a communications discipline rather than a search one, and it is the thing most teams get backwards on the first attempt.

What AI reputation management actually is

Strip out the vendor language and there are three jobs.

The first is finding out what is being said. Not in the abstract, but in the specific words a buyer would use, on the engines your buyers actually use, recorded in a way you can compare against next month.

The second is judging it. An answer can be positive and wrong. It can be accurate and still lose you the deal because it names a competitor in the last line. Sentiment alone is a thin read here.

The third is changing it, which is slower and less glamorous than the first two, and where most programmes quietly stop.

None of this is reputation management in the old sense. Old-school online reputation management was about pushing a bad result down a page. There is no page here. There is one answer, written fresh each time, and either you are in it or you are not.

Why this landed on the comms desk

The instinct is to hand it to whoever owns SEO, and the instinct is understandable. Both involve search boxes. Both involve content.

But look at what the two disciplines are actually competing for.

SEO competes for a slot. Ten blue links, and the work is to occupy a better one. The unit of value is a page, and the page is yours, on your domain, under your control.

AI reputation management competes to be quoted. There is no slot. A model writes a sentence about your category and decides, somewhere inside that process, which claims to make and which sources to lean on. The unit of value is not a page. It is a sentence about you that exists somewhere on the internet, and the overwhelming majority of those sentences are on property you do not own.

Your website is one input into what a model says about you, and usually not the most trusted one. The rest is journalism, reviews, forums, filings, directories and other people’s opinions, all of which are comms problems.

That is why this belongs to communications. The failure modes are reputational, the raw material is earned coverage and third-party credibility, and the fixes are editorial before they are technical. Search teams are essential to the execution. They are the wrong owner of the outcome.

The four ways a model gets your brand wrong

In practice, problems fall into four buckets, and they need different responses.

Stale facts. The leadership is out of date, the funding round is three rounds old, the product line is one you retired, the ownership structure changed and the answer did not. This is the most common and the most fixable.

Mistaken identity. Another company shares part of your name, or operates in an adjacent category, and the model has blended you. The answers read as plausible nonsense. This one is badly under-diagnosed because the text looks fine until a subject-matter expert reads it.

Inherited framing. One critical article, one regulatory story, one unusually well-optimised complaint thread has become the summary. Everything the model says is technically sourced and the overall impression is wrong. This is the hardest of the four, because there is nothing false to correct.

Absence. You are simply not there. Ask for the leading companies in your category and you are not named. No error to point at, no negative sentiment to flag, just a recommendation that goes to someone else.

Absence is the expensive one and it is the one almost nobody measures, because monitoring tools are built to find mentions and a non-mention is not a mention. If your programme only reports on what was said about you, it is blind to the failure mode that costs the most.

Where an answer actually comes from

You cannot fix what you cannot trace, so it helps to be precise about the machinery.

An assistant’s answer is assembled from two sources, in a mix that changes by engine and by question.

The first is what the model absorbed during training. Call it memory. It is broad, it is not attributable, and it moves on the slow clock of model releases. You influence it over quarters and years, by what exists about you at scale, and you will never get a receipt for the influence.

The second is retrieval: the engine runs a search, reads a handful of live pages, and writes an answer grounded in them. Perplexity and Google’s AI Overviews lean heavily this way and will show you their sources. ChatGPT, Gemini and Claude do it for some questions and not others.

The practical consequence is the most useful thing in this article:

The retrieval path is the one you can move this quarter. The memory path is the one you move over years. Any plan that does not separate the two is going to promise timelines it cannot keep.

When an engine names its sources, you have been handed the source chain for free. Read those pages. That list is the actual brief for the next three months of comms work, and it is a far better brief than a keyword report, because it tells you precisely which third parties are currently speaking on your behalf.

What you can control, what you can influence, what you cannot

Being honest about this is not modesty. It is what keeps a programme from being judged against a promise nobody could keep.

You control what your own properties say: the site, the newsroom, the documentation, the pricing and policy pages, the structured data, and whether any of it is written in a form that can be lifted cleanly into an answer.

You influence what third parties say: press coverage, analyst and review platforms, Wikipedia and comparable reference sources, industry directories, community threads. Influence here means the ordinary work of communications, done with an awareness that the audience now includes a machine that will paraphrase the result.

You do not control the model, its weights, its ranking of sources, or its output on any given day. Nobody does. Any vendor who offers to remove something an AI says about you is selling a capability that does not exist, and that claim alone should end the meeting.

How to audit what AI says about your brand

Do this properly once and you will never accept a screenshot as evidence again.

Write prompts the way a buyer writes them. Not your brand name on its own. Real intent, in natural language, across five types:

1. Category discovery. “Who are the leading providers of X for mid-sized manufacturers?” This is where absence shows up.
2. Direct brand. “What is [brand] and who owns it?” This is where stale facts and mistaken identity show up.
3. Comparison. “[Brand] vs [competitor], which is better for X?” This is where inherited framing shows up.
4. Objection and risk. “Is [brand] reliable? Has [brand] had any controversies?” Ask it. If you will not, your buyer certainly will.
5. Procurement. “What should I ask [brand] before signing?” The answers are a free read on how the market frames your weaknesses.

Run each prompt on every engine that matters to your market, which today means ChatGPT, Gemini, Claude, Perplexity and AI Overviews at minimum.

Record four things every time: the engine, the date, the answer in full, and every source it named. The sources matter more than the answer. The answer is the symptom.

Then do it again next month, unchanged. Answers drift. A model update, a new article, a competitor’s product launch, and the paragraph is different. One reading is an anecdote. A series is a measurement. This is the whole reason AI reputation tracking is built as a daily check rather than a report you commission twice a year.

A note on scale. Five prompt types across five engines is twenty-five readings, and a serious audit uses far more prompts than five. Doing it by hand is genuinely useful the first time, because you will read the answers properly, and it stops being viable somewhere in the second month.

The repair work, in the order that works

Repair is source work. Every step below is aimed at the chain, not at the answer.

Trace before you write. Take the specific claim you want changed and find the pages carrying it. When the engine cites, you have them. When it does not, search the claim in the model’s own phrasing and the sources usually surface fast.

Fix your own properties first, because it is the only step entirely within your control and because retrieval favours pages that answer a question directly. One question per page. The answer in the opening lines, not in the eleventh paragraph. Claims that are specific, checkable and dated. A model summarising a vague page produces a vague sentence about you.

Get the fact corroborated somewhere that is not yours. A correction that appears only on your own site is a claim. The same correction in trade press, a reference source or an analyst note is a fact, and models weight it accordingly. This is ordinary media relations, which is exactly the point: the discipline is not new, only the audience is.

Make the corrected page quotable. This is where generative engine optimization genuinely helps, and it is narrower than it sounds. Clear headings that match real questions, self-contained paragraphs that survive being lifted out of context, explicit dates, plain statements of fact rather than marketing compression, and structured data that says what the page is. None of it will rescue a page with nothing to say.

Re-test, on a schedule, and write down the result. Changes take weeks to appear in retrieval answers and longer in memory. A programme with no before-and-after record cannot tell the difference between a fix that worked and a model update that happened to help, which means it cannot repeat either.

The sequencing matters because the temptation is always to jump to step four. Publishing quotable content about a claim you have not traced is how teams spend a quarter optimising a page that was never in the source chain.

Measuring it without inventing a number

The category is already filling up with composite “AI visibility scores”, and you should be suspicious of all of them. A single number here is unfalsifiable by construction, because there is no ranking, no impression count and no click to check it against.

Four things can be measured honestly, and together they say plenty.

Measure	The question it answers
Presence rate	In what share of category prompts are we named at all?
Answer sentiment	When we are named, how are we characterised?
Factual accuracy	What proportion of claims about us are correct and current?
Recommendation rate	When the assistant picks one, how often is it us, and who wins otherwise?

Report them as four numbers, with the prompt set and the date attached, and never average them into one. Presence rate and factual accuracy fail for completely different reasons and are repaired by completely different work. A composite hides exactly the information the programme needs.

One more discipline worth borrowing from the rest of reputation measurement: state your sample. A recommendation rate calculated from twelve prompts is not a percentage, it is a handful of anecdotes wearing a percent sign. If the sample is thin, say so on the slide.

Who owns this, in practice

Communications owns the outcome. The failures are reputational, the raw material is earned credibility, and the judgement calls about what to correct and what to leave alone are comms calls.

Search and web teams own a large slice of the execution, and the collaboration is genuinely productive as long as the objective is set correctly. Ask an SEO team for visibility and you will get visibility work. Ask for accuracy and citation and you will get something far more useful.

Legal comes in when an inaccuracy is material, particularly for listed companies, regulated categories and anything touching safety. A model asserting something false about a medicine, a financial product or a public service is not a marketing problem.

And somebody needs to own the schedule, because the single most common way this work fails is not being wrong. It is being done once, in a burst of enthusiasm, and never repeated.

Where this goes next

Two things are already visible.

Assistants are becoming the first stop rather than the second, which means the answer increasingly is the impression, with no page visit behind it to correct the record. The gap between what people believe about you and what your analytics can see is going to widen, and monitoring the answer layer is the only instrument that closes it.

At the same time, engines are leaning harder on retrieval and citing more openly. That is good news, because a cited source chain is an actionable brief. The brands that treat those citations as a work queue rather than a curiosity will pull ahead of the ones still waiting for a best-practices deck.

The work itself is not exotic. Find out what is being said, trace where it comes from, correct it at the source, prove it elsewhere, check again. Communications teams have done exactly that for decades. The only real change is that one of the outlets now writes a fresh article every time somebody asks, and never sends you a clipping.

If you want to see this running against live data rather than read the argument, that is what the AI reputation module does, and what the Action Center turns the findings into.

Common questions

What is AI reputation management?

AI reputation management is the practice of tracking what AI assistants such as ChatGPT, Gemini, Claude, Perplexity and Google's AI Overviews say about a brand, and correcting it. Because no one can edit a model's output directly, the work is done on the sources those models read and cite.

How is AI reputation management different from SEO?

SEO competes for a position on a results page. AI reputation management competes to be the source a model quotes when it writes an answer, and there is no position to win. The tactics overlap on the technical side, but the goal is citation and accuracy rather than ranking, and the owner is usually communications rather than search.

Can you remove false information an AI says about your brand?

Not directly, and any vendor promising to delete a model's output is describing something that does not exist. What works is finding the sources feeding the claim, publishing a clear and dated correction where models retrieve from, earning third-party corroboration, then re-testing the prompt over several weeks to confirm the answer has moved.

How often should you check what AI says about your brand?

Monthly is the minimum for a stable category and weekly is sensible for anything in a live news cycle, because retrieval-based answers can change within days of new coverage. A single screenshot is an anecdote. The value is in running the same prompts on a schedule so you can see drift.

Who should own AI reputation management inside a company?

Communications should own it, because the failures are reputational and the fixes are editorial. Search and web teams execute the technical half, and legal is brought in when an inaccuracy is material. Leaving it with SEO alone tends to produce visibility work with no view of what is being said.