An outdated address, an old price, a service you don’t even offer. You can’t delete it. We find the source the statement comes from, correct it and monitor what the models say afterwards.
The obvious question is: where do I report this? The answer is unsatisfying. None of the major providers offers a correction form for brands.
A language model has no database containing your company entry. It repeats what it learned from the web or just found there. A false statement is therefore rarely a model error. It usually reflects what is written about you somewhere.
That points to the way forward: don’t fight the answer, change the source it comes from.
In this order, because each step builds on the previous one.
First we clarify where the statement comes from. With services that search the web live, the sources are listed below the answer. For statements from the training data, we look for the place on the web that matches best. Without this step you only treat the symptom.
The correct information goes on a page you own, with a date and phrased so a model recognizes it as a statement. Not hidden in body text, but as a clear sentence. This is the part you control yourself.
Directories, industry portals, press articles, sources close to Wikipedia: where the false information appears, we contact the operator. Where the correct information is missing, we make sure it appears there. This is the most time-consuming part and the one with the biggest effect.
A prompt set about your brand runs regularly against ChatGPT, Perplexity, Claude and Google AI Overviews. You see whether the correction lands, and a new false statement shows up before a customer reports it.
Google accepts feedback on the AI Overview, Perplexity through the flag icon below the answer. We use these channels, but we don’t rely on them: they are non-binding, have no deadline and give no notice of the outcome.
For statements that damage your business, Google runs a separate legal reporting channel. We tell you when we see a case in this category. A lawyer then represents you, not us.
Nobody can forbid a language model from saying a sentence. Anyone who promises that is selling you something they can’t deliver.
What we influence is the basis. Once the sources are right, the answer is very likely to change the next time the system processes them. When that happens is up to the provider.
With services that search live, this can take effect within days. If the statement comes from a model’s training data, it can take until the next model update. That is why we don’t name a date.
We query the systems with a prompt set about your brand and record what they say. This produces the list of incorrect statements, sorted by potential damage.
For each false statement: find the source, correct your own pages, update third-party entries. Anything held by third parties takes longer, because someone else decides there.
The same prompt set keeps running. In the monthly report you see which statements have changed and which still stand. Whatever persists, we tackle again.
This work requires knowing which questions your brand appears in at all. Prompt mapping clarifies that. It also needs content that a model finds useful as a source, which is what citation engineering takes care of.
Our GEO and LLMO service describes the bigger picture. If you want to start on your own, you will find the steps in the Swiss GEO framework (PDF in German).
If AI answers attribute your brands incorrectly because several names and domains coexist after an acquisition, the work starts earlier: with your brand architecture.
Brand sentiment is part of our GEO service, from CHF 1,800 / month, cancel monthly. 12 months: from CHF 1,710 / month (5% discount).
Not as a company. OpenAI’s privacy portal accepts requests about the personal data of natural persons, requires an official ID and weighs every case against freedom of expression and the public interest. A false statement about a company does not fall under this. None of the major providers offers a correction form for brands.
Google accepts feedback on an AI Overview with a category and free text, Perplexity through the flag icon below the answer. Neither is binding: there is no deadline, no notice of the outcome and no right to a correction. We use these channels, but we do not base the work on them.
Usually for one of three reasons. The information was once on the web in that form and is now outdated. Or it sits on a page we do not control, such as a directory. Or the model combines two correct pieces of information into a false statement. Tracing the source shows which case applies.
That depends on when each system re-reads its sources. With services that search the web live, it can take days. For information that comes from the training data, it can take months or last until the next model update. No point in time can be guaranteed.
Then it belongs with a lawyer, not an agency. Google has its own legal reporting channel for such cases, which also covers knowledge panel cards. We tell you when we see a case in this category, but we do not provide legal representation.
Brand sentiment work is part of our GEO and LLMO service, which you can cancel monthly. You will find the price in the section above.
We query the systems with a prompt set about your brand and show you which statements are wrong. Free and with no obligation. You can work with us in English.