Six engines have an opinion about you. None of them has an appeals process.

The reputation layer nobody has a process for

A buyer used to check you by searching. They saw ten results, formed their own view, and you had levers on most of that page. You could outrank a bad result, ask a platform to remove something, or reply to the review in public where the reply was part of the record.

Now a large share of that checking happens inside a conversation. The buyer asks an assistant whether you are any good, and gets one paragraph back. There is no page of ten results. There is one answer, delivered privately, and you never learn it was given.

That answer is assembled at the moment it is asked, from sources the engine chose. Ask the same question tomorrow and the wording changes. Ask it from a different country and the sources change. There is no ranking to climb and no position to hold, which is why the old playbook does not transfer cleanly.

Three things you cannot do, and one you can

It is worth being blunt about the limits before describing the work, because most of what agencies sell in this category quietly assumes powers that do not exist.

There is no takedown form. No engine offers a mechanism for a business or an individual to have a specific claim about them removed from generated answers. Support routes exist for other things. This is not one of them.

There is no appeal. Nothing reviews an answer for accuracy on request. Some engines let you report a response, which feeds general model improvement rather than your case.

The legal route is currently closed. The first defamation claim over an AI-generated statement to reach judgment, Walters v. OpenAI in Georgia, ended in summary judgment for OpenAI in May 2025. The court accepted that a reasonable reader would not treat the output as a statement of fact, and that disclaimers plus the absence of proven negligence defeated the claim. Other cases are moving through courts elsewhere and the position may change. Today it is not a route to plan around.

What is left is the supply side. Engines do not invent from nothing. They retrieve, and they retrieve from a knowable set of places. Change what is available to be retrieved and the answer changes, because the answer was always a summary of what was there.

Where the answers actually come from

The published evidence disagrees with itself, and we would rather show you that than pick the number that flatters the pitch.

Yext analyzed 6.8 million AI citations and found roughly 86 percent traced back to first-party pages and business listings, with community forums at around 2 percent. Several independent studies of AI Overviews and ChatGPT put Reddit far higher, near 40 percent of cited sources in some category sets. The gap is real and it is explained by question type. Ask about a named company and the engine goes to owned properties and structured listings. Ask an open recommendation question with no company named, and it goes to forums and roundups.

That distinction is the whole strategy. Both questions get asked about you, and they need different work.

Two findings hold across the studies. Review platforms carry weight out of proportion to their traffic, with profiles on Trustpilot, G2 and Capterra roughly three times more likely to be cited than an equivalent page elsewhere. And engines rank their sources informally by perceived reliability, which is why an encyclopedia entry or an analyst note outweighs a dozen enthusiastic blog posts.

Four ways an answer goes wrong

In practice the failures we find fall into four patterns, and each one has a different fix.

Stale fact. The engine reports something that was true. A closed matter reads as open, a former partner reads as current, a price from two years ago reads as this year’s. This is the most common and the most fixable.

Merged identity. Your record fuses with somebody who shares your name. Common for individuals, and for companies operating in more than one market under similar names. Nothing about you is wrong except that it is not you.

Invented specifics. The answer contains a detail with no source at all: a case that never happened, a certification never held, a figure nobody published. Harder to fix, because there is no wrong page to correct. The fix is to make the correct fact easier to find than the gap that invited the invention.

Omission. The answer is not wrong about you. It simply does not mention you, and names a competitor instead. Commercially this is often the most expensive of the four, and it is the one that never shows up in a monitoring alert because there is nothing to alert on.

How we work on it

The engagement has four parts and they run in order, because each one depends on the last.

Establish the baseline. We agree a prompt set with you, normally between forty and eighty questions, written the way your buyers actually ask rather than the way your marketing describes you. It covers the direct question about your name, the open recommendation question where you are not named, the comparison against your two or three real alternatives, and the awkward question your sales team already dreads. We run it across six engines, in the languages your market uses, and store the full answer text with its date. That archive is the only evidence base that exists, because nothing is retained on the engine side. It lives in your client portal, where every run is stored with its date and you can re-read an answer from three months ago instead of trusting a screenshot.

Trace the sources. Every problem answer gets traced back to what produced it. Perplexity makes this easy because it numbers its citations. The others require comparison work: run the question with retrieval on and off, vary the phrasing, and see which sentences survive. What comes back is a specific list of pages, listings and threads rather than a general sense that the internet is unkind.

Fix the supply. This is where the time goes, and what it involves depends entirely on what the tracing found. Correcting and dating your own pages so the current fact is unambiguous. Repairing structured data so the machine-readable version of you matches the human one. Getting the entity right across the places engines treat as authoritative, which for most clients means Wikidata, the relevant industry registers, and the review platforms. Building the specific page that answers the question no existing source answers well, which is how you become the source rather than the subject. Where a genuine grievance sits behind a bad thread, we will tell you that the fix is operational and not editorial.

Re-run and report. The same prompt set, monthly, on the same schedule. Answer changed, unchanged, or worse, per engine, with the previous month’s text beside the current one. Perplexity tends to move first because it retrieves for nearly every query. Anything held in a model’s training memory rather than its index waits for a release, and we will say so instead of describing a delay as progress. The month-on-month comparison sits in the portal alongside your search and paid media reporting, on read-only connections.

What we will not tell you

We will not promise a specific answer by a specific date, because no one controls the generation step. We will not sell a single visibility score, because the six engines disagree and an average hides the one that is failing. We will not describe review generation or thread manipulation as reputation work. And we will not take the engagement if the honest finding is that your operations produced the complaints, because no amount of source work outruns a real problem for long.

What we will do is show you exactly what six engines say about you today, explain which sources produced it, fix the ones that can be fixed, and re-run the set every month so the direction of travel is a matter of record rather than opinion.