AI Reputation Radar
Correct the record at source, so AI says what is true about your organisation.
We build a fact file with you, the version of your organisation you would stand behind, then sweep ChatGPT, Gemini, Perplexity and Claude and check every answer against it. You get each wrong statement logged with the engine that made it, the source it came from, and the correction that fixes it at source.
Senior-led Checked against your own facts Every error traced to a source
The fact file
- Legal name, trading names and registration
- What the organisation does, and for whom
- Leadership, with roles and the dates they applied from
- Ownership, group structure and named investors
- Locations, operating markets and the regulator that applies
- Financial position, exactly as you state it
- Credentials, certifications, memberships and awards
- The current status of anything contested
Approved by you before the first sweep runs
The deliverable
A fact file built line by line with you. Every engine, checked against it.
Nothing on this page compares an engine's answer with our opinion of your organisation. It compares it with a document you wrote and signed off: the version of the facts you would correct a journalist on.
That document is the benchmark for every sweep that follows, and it is the reason a finding can be argued about. An error is a departure from your record, not from our reading of you.
Eight classes of fact go into it, and each one carries its own score in the log, so a weak area cannot hide inside a comfortable average.
From fact file to correction plan, in three stages.
A tool can tell you the answers differ. Deciding whether a difference is an error, an omission or an unflattering truth is a judgement, and a person makes it.
-
Stage 01
We build the fact file
A short working session and a document: the legal name, what you do, who runs you and since when, ownership, jurisdictions, the numbers as you state them, the credentials you actually hold, and the current status of anything contested. You approve every line. From that point on the file is the benchmark, so the exercise measures the engines against your record rather than against our impression of it.
-
Stage 02
We sweep the engines
The same question set goes to ChatGPT, Gemini, Perplexity, Claude and Google's AI Overviews, and every answer is captured word for word and dated. Then each statement is checked against the file, line by line. A tool can tell you the answers differ. Deciding whether a difference is an error, an omission or simply an unflattering truth is a judgement, and a person makes it.
-
Stage 03
You get the log and the plan
Every wrong statement, quoted, with the engine, the date, the evidence and the source that appears to be producing it. Then the corrections in priority order, with an owner on each: the ones you can make this week, the ones your advisers make, and the third-party sources that need an approach rather than an edit. The next sweep opens by marking off what moved.
Every wrong statement, quoted, with the source behind it.
The log is the deliverable. Each line is a statement an engine made about you, the engine that made it, whether it matches your file, and the material that appears to have produced it. Statements that match are logged too, because the ones that hold are as much a finding as the ones that do not.
| What the engine said | Engine | Against your file | Where it comes from |
|---|---|---|---|
| Founded in 2009 | Perplexity | Wrong | A trade directory entry |
| Led by its founder as chief executive | ChatGPT | Out of date | An undated team page |
| A subsidiary of the group | Gemini | Superseded | An archived group site |
| Operates across three named markets | Claude | Matches | Your own site |
| Holds the standard it lists | ChatGPT | Matches | Your own site |
| Facing proceedings, according to reports | AI Overviews | Resolved, not stated | The original coverage |
Sample data throughout. Yourfirm.com is not a client.
Inside the document // six parts
-
The accuracy score and the spread
One number for how much of your record the engines currently state correctly, out of 100, built as the exact average of the classes beneath it. Beside it, the spread: which parts of your record hold and which are drifting.
-
The discrepancy log
Every statement that contradicts your fact file, quoted in full, with the engine that made it and the date it was captured. Errors are separated from findings that are accurate but unflattering, because those two things need completely different responses.
-
The evidence
The captured answer behind every entry, stored so it can be produced later. When a correction is disputed, or when a colleague asks whether it really said that, the log has the answer as it stood on the day, not a summary of it.
-
Where the error comes from
For each discrepancy, the source that appears to be producing it: a stale profile, a superseded filing, an old article, a page of your own that never stated the position plainly. This is the part that turns the log from a grievance into a task list.
-
The correction plan
Every discrepancy restated as an action, ordered by what it costs you to leave in place, with an owner on each line. Your team for the pages and statements you control, your advisers for filings and registrations, us for the third-party sources that need an approach rather than an edit.
-
The watch list
What to expect at the next sweep: which corrections should have worked through, which are slower because the source is widely republished, and which facts are worth watching because they are about to change.
Errors about an organisation are not random.
They fall into a small number of recognisable classes. Each one fails in its own way, arrives by its own route, and gets corrected in its own way. Every quotation below is a specimen of the class, not a claim about anyone.
-
Founded in 2009, the company is a regional distributor of pumping equipment.
Identity and activity
The summary sentence sits underneath every other answer, so an error here propagates into all of them. We check the description each engine gives against your own: the activity, the markets, the model, the stage. The common failure is not a falsehood but a description that was true three years ago and has been quietly repeated ever since.
How it gets in
- an old about page
- a trade directory
- a funding announcement
-
The company is led by its founder, who continues to serve as chief executive.
Leadership, roles and dates
Named people are where engines are most confident and most often wrong. Directors who left are still listed, titles are inflated or demoted, and a biography is merged with a namesake who happens to share the name and the sector. We check every named individual, their stated role and the date it applied from.
How it gets in
- an undated team page
- a conference biography
- a namesake profile
-
It has raised approximately the amount reported in its last announced round.
Numbers and dates
Funding, headcount, revenue, founding year, contract values. Numbers travel further than anything else and correct themselves less often, because a figure repeated by a data aggregator reads as current to every engine that consults it. We check each stated number against your record and trace it back to whoever is still publishing the old one.
How it gets in
- a data aggregator
- an old press release
- a sector round-up
-
It operates as a subsidiary of the group that divested it.
Ownership and structure
Parent companies, subsidiaries, joint ventures, investors, group relationships. Engines flatten corporate structures readily, attributing a subsidiary's work to the parent or leaving a divested business attached to you years later. In regulated sectors this is not a cosmetic error, it is a compliance problem in someone else's search results.
How it gets in
- a superseded filing
- an archived group site
- a deal database
-
Headquartered in the city it moved out of, and regulated accordingly.
Location and jurisdiction
Where you are registered, where you operate, which regulator applies. We check the geography every engine states, because a company placed in the wrong jurisdiction is answered against the wrong rules, and because an operating footprint that has moved on is one of the least-corrected facts on the internet.
How it gets in
- a stale registry entry
- an old contact page
- a listings site
-
The company holds a certification it has never applied for.
Claims, credentials and certifications
Accreditations, memberships, standards held, awards, published commitments. Two failures matter here in opposite directions: credentials you hold that no engine mentions, and credentials attributed to you that you have never held. The second is the one that ends up in a tender response written by somebody else.
How it gets in
- a supplier listing
- an award long-list
- a partner's website
-
The company is facing proceedings, according to reports.
Contested matters, and how current they are
Litigation, regulatory action, disputes, criticism. The finding is rarely that an engine invented something. It is that a resolved matter is still described in the present tense, with no mention of the outcome, because the coverage of a resolution is always thinner than the coverage of the allegation. We check status, framing and date on every one.
How it gets in
- the original coverage
- an undated summary
- no record of the outcome
-
Its flagship project was delivered in a market it has never worked in.
Attribution and conflation
Whether the answer is actually about you. Engines merge organisations with similar names, attribute a competitor's project to you, or fold a former group company's history into yours. This class produces the most damaging errors in the log, because everything in the answer is true, and none of it is yours.
How it gets in
- a name collision
- a shared trading name
- a former group company
How corrections work
You cannot edit an engine. You can edit what it reads.
Every wrong answer is downstream of something real: a profile nobody retired, a filing that was superseded, a page that never stated the position plainly. Correct that source and you change what the engine says and what the next journalist finds.
This is why the log stops at a source rather than at a screenshot. There is no complaints process for an AI answer and no editor to write to. There is only the material the engine is reading, and most of it is ordinary communications work: a page with dates on it, a directory entry brought up to date, a resolution stated as plainly as the allegation was. Do that, and the correction reaches the engine and the reporter by the same route.
An error repeats itself.
Engines learn from what is published, and what is published now includes text engines helped write. A wrong figure gets quoted, summarised and republished until it stops looking like anybody's mistake and starts looking like the record.
None of this is adversarial and nobody is watching it happen. A director who left, a round that was superseded, a case that was resolved: each is a small clerical fact, and correcting the source changes what a funder, a counterparty or a journalist finds next.
Corrections take time to work through the sources engines read, which means the useful moment to start is a quiet one. The week you need the answer to be right is the week it is too late to begin fixing it.
- ChatGPTA subsidiary of the group.
- GeminiPart of the group, operating as a subsidiary.
- PerplexityA group subsidiary, according to available sources.
All three are reading one archived page
Three engines, three phrasings, one source. Correct the page and the log tells you which of the three moved.
What is AI Reputation Radar?
It is a standing accuracy check on what AI engines say about your organisation. We build a fact file with you, the verified version of who you are, who runs you, what you have done and what you have not, then sweep ChatGPT, Gemini, Perplexity, Claude and Google's AI Overviews and compare every answer against it. What comes back is a discrepancy log: each wrong statement, the engine that made it, the source behind it and how to correct it.
How do I correct wrong information about my company in AI answers?
By correcting the material the engine is reading, because that is the only lever anyone actually has. There is no form to submit and no editor to write to. In practice a wrong answer traces back to a small number of sources: a directory profile carrying an old leadership list, a superseded filing, an article nobody updated, or a page of your own that never stated the current position plainly. Correct those, then measure whether the answers moved. That measurement is the part most organisations skip and the reason they cannot tell whether anything worked.
Why do AI engines get things wrong about us?
Rarely out of malice, and almost never out of nowhere. An engine assembles an answer from whatever it can reach and trust, so an old directory profile, a press release nobody retired, a filing that was superseded or a namesake company with a similar registration can all end up stated as current fact. The engine has no way of knowing which of its sources is the one you would stand behind. The fact file is how it finds out.
What goes into the fact file?
The things you would correct a journalist on: legal name and trading names, what the organisation does and for whom, leadership with roles and dates, ownership and structure, locations and jurisdictions, funding or financial position as you state it, credentials, certifications and memberships, and the resolved status of anything contested. You approve every line before it becomes the benchmark, because the whole exercise measures answers against your record, not ours.
Who inside an organisation owns this?
Usually communications, often with the general counsel and whoever owns risk in the room, because the findings land in three different in-trays. A wrong description is a communications problem. A wrong statement about ownership, jurisdiction or a contested matter is a legal and compliance one. A wrong credential is both. The correction plan is split by owner for exactly this reason, so nobody receives a list that is mostly somebody else's work.
How often do you sweep?
Monthly by default, and on demand around anything that changes the record: a leadership change, a funding event, a restructure, a piece of coverage you would rather did not become the summary. Frequency is agreed with you at the start. What matters more than cadence is that the question set and the fact file stay constant between sweeps, so a change in the log is a change in the world rather than a change in the method.
What do you do about an error once you have found it?
The log names the source, not just the symptom, because correcting the source is the only thing that changes the answer. Some corrections are yours to make in an afternoon: a page that never stated the current position plainly, a leadership list with no dates on it. Others need an approach to a third party, a data aggregator or a publisher. We mark which is which, and we do the ones that are communications work.
Can you guarantee the engines will stop saying it?
No, and anyone who says otherwise is selling something. No one can edit an engine's output directly. What can be changed is the material the engine reads, and in practice a well-sourced correction to a widely-read profile does work through into answers, sometimes in weeks. The Radar's job is to make that measurable: the next sweep shows you whether the correction landed, on which engines, and where it did not.
How is this different from media monitoring?
Media monitoring tells you what was published. The Radar tells you what is being said, right now, in the place most people now ask first, and whether it is true. An AI answer is not a clipping. It is a summary written fresh each time from sources you may never have seen, and it carries no byline to hold to account. Monitoring finds the article. The Radar finds the sentence an engine built from it three years later.
Is this the same as the AI Visibility Dashboard?
No, and they answer different questions. The Dashboard asks whether you are present and how you compare: are you named, are you cited, who else is in the answer. The Radar asks whether what is said is true. A company can be highly visible and consistently misdescribed, which is the worst of both. Clients running both use the Dashboard for position and the Radar for accuracy.
Request the radar
Make your own record the one AI repeats.
Tell us who you are and where to reach you. We come back with the fact file to approve, and the first sweep follows with every discrepancy, its source and its correction.
Fact file approved before anything runs
