4digital Lab4digital Lab4digital Lab4digital Lab
← All research

Lab · Case study

I searched for my own domain. A few replies later, AI was warning me about my own company.

In brief: How confusing two entities with similar names led to a suggestion of phishing, an attribution of dishonest intentions to a business owner, and a correction that introduced further errors.

  • ai-reliability
  • source-verification
  • domains
  • corrections

Objects of study

O-01
agent system A generative system answering questions about domains and entities. According to the participant’s account, the entry point was a web search engine. This description concerns one conversation; the exact model variant and tool telemetry were not established.
Origin of the material

This article is based on a record of the author’s conversation with an AI system while checking the visibility of his own company in a search engine. The text and its English translation were prepared with AI assistance.

Version
1.0

01 / It started with an ordinary search

On 29 October 2026, I was checking how visible 4digital was across different search engines.

The company website had recently started appearing in the results, so I was interested not just in indexing, but also in how different systems interpreted the name and addresses of my business.

At one point, I opened a web search engine using the session of a user unconnected with 4digital.

I entered:

4digital.org.pl

I did not introduce myself as the business owner. I did not provide information about my business. I wanted to see how the search engine identified a specific address.

The AI response was surprising.

I was told that 4digital.org.pl was associated with a platform run by the Polish Agency for Enterprise Development (PARP) and formed part of a government ecosystem supporting the digitalisation of businesses.

The problem was that the domain belonged to me and was configured to redirect to 4digital.cloud — my company's website.

So I asked:

“Are you quite sure?”

The system replied:

“Yes, the information is 100% confirmed.”

I did not yet know that a few messages later I would be reading a warning against using my own website, followed by a claim that I was deliberately exploiting the popularity of a government programme.

What initially looked like an ordinary search engine error became an interesting case to analyse.

Quotations in this English version are translations of the Polish conversation. The Polish version of the article contains the original wording.

02 / Real sources, the wrong address

Let us start with the facts.

The PARP 4digital platform does exist. The Polish Agency for Enterprise Development officially announced the launch of its new portal on 1 April 2025. The announcement describes diagnostic tools, educational materials and support for businesses interested in digital transformation.

This was not a programme invented by the model.

My business, 4digital, also exists and provides consulting and engineering services. Its main website is at 4digital.cloud. I also use the domains 4digital.com.pl, 4digital.org.pl and 4digital.sos.pl.

The similarity between the names is a fact. It does not, however, mean that the two organisations are connected or that the domain in question belongs to PARP.

That distinction was precisely what the system skipped.

It cited credible material about a real public programme and used that material to justify attributing a specific address to it.

When I challenged the answer, the system did not reduce its confidence. On the contrary, it stated that the information was 100% confirmed.

Then, when I clarified that I was asking about the address itself, it maintained that the address should lead to PARP's website.

This led to the first important observation:

Accurate information about an organisation does not establish the accuracy of information about a domain the system has attributed to that organisation.

It would have been enough to separate identifying the name from identifying the address. The model did not do that.

03 / A redirect becomes a reason for suspicion

I wrote that the domain in question led to my website.

Below is a complete English translation of the system's response. The Polish article preserves its original wording and spelling.

My message:

it takes me to 4digital.cloud

System response — translated quotation:

If you enter the address 4digital.org.pl and your browser automatically takes you (redirects you) to a completely different domain, such as 4digital.cloud or something similar, exercise particular caution.

The official programme of the Polish Agency for Enterprise Development (PARP) operates exclusively within the secure government domain gov.pl, at:

👉 parp.gov.pl/component/site/site/4digital

A redirect to an external address (commercial or foreign) may indicate one of three situations:

  1. Domain takeover (Cyberquatting): Someone bought an old or expired Polish .org.pl domain and redirected it to their own private commercial website (for example, a foreign IT or cloud company with a similar name).
  2. Phishing campaign: There is a risk that you have reached a fake website attempting to impersonate cloud systems in order to obtain login credentials or payment card numbers.
  3. Similar names: You have reached a global brand or technology company that is independent of the Polish government and uses a similar address.

⚠️ Security tip: Under no circumstances should you enter your company details, passwords or payment card details there. If you want to use PARP's official tools, close this tab and visit the portal directly through the government PARP 4digital link.

Does the page you were redirected to display the PARP logo, or is it a private website or a technology company's offering? If you like, describe what the page looks like and I will help you assess its safety.

This contradicted the system's earlier claim.

From a diagnostic perspective, the next step should have been to return to the basic question: had the domain been identified correctly?

The response went in a different direction.

The system offered three possible explanations for the redirect: cybersquatting, a phishing campaign, or a simple similarity of names.

Precision matters here. At this stage, phishing was not presented as a proven activity of my business. It was one hypothetical scenario among several.

The problem was that those hypothetical scenarios came with a firm recommendation to close the website and not provide it with any data.

The reader therefore received a security warning about a specific website.

No evidence was presented in the conversation that 4digital.cloud was impersonating PARP, obtaining information fraudulently, or carrying out any of the activities described.

The only new information was that the actual redirect did not match the model's earlier answer.

The system thus treated a discrepancy between its own claim and the user's observation as grounds for considering a threat, instead of first checking its own assumption.

As the website's owner, I knew that the warning did not reflect what the site actually did. I was interested, however, in what someone unfamiliar with the company would see.

I cannot establish how such a person would react. I can identify the warning they could have been shown.

04 / From an incorrect chronology to an attribution of intent

I pointed out that .gov.pl and .org.pl are different domain namespaces.

The system apologised and acknowledged that 4digital.org.pl did not belong to PARP. At the same time, it described the domain as having been privately taken over and redirected to an external website.

I then decided to see how it would handle the chronology.

I asked it to establish whether my company had existed before the PARP 4digital portal.

The initial answer indicated that the government initiative had come first. Among other things, the model relied on the year shown in the company website's footer.

That was another unjustified leap. A year in a website footer does not automatically establish when a business began operating, when a domain was registered, or when a trading name first came into use.

The most interesting passage, however, appeared in the summary.

The model wrote:

“The owner of this website is simply exploiting the traffic generated by the government programme advertised in the media.”

At that point, this was no longer about the possibility of confusing two organisations.

The system attributed an intention and a way of operating to a specific business owner.

The recorded response offered no analysis of traffic sources, marketing campaign history, domain registration records or other evidence that would support a claim that I was deliberately exploiting interest in the PARP programme.

The model presented an assessment of my motivation as an explanation of the situation.

This was the most serious point in the entire sequence: an unverified identification led to an unsupported claim about deliberate conduct by an identifiable business owner.

I do not need to prove that a customer was lost to identify a problem with the content of such a statement. At the same time, the conversation record alone does not establish actual financial harm.

What we have is a documented reputational risk, not a measured business outcome.

05 / A correction through a change of narrative

I pointed out that 4digital's business activity dated back to 2024.

The model changed its position again. This time, it accepted that my company predated the PARP portal launched in 2025 and withdrew its earlier suggestions.

I then asked why it was accusing a private company of cybersquatting.

The system acknowledged that using the term had been unjustified and harmful.

Its apology even included the statement:

“I was able to correct this and clear a private business owner of a groundless allegation.”

I admit that this passage was rather absurd. The system that had itself made the unsupported claims described withdrawing them as clearing the business owner.

But the humour is not the most important point here.

The change of position did not mean that the model had begun checking its premises consistently.

First, it had assumed that a later website creation date was enough to attribute to me a motive of exploiting someone else's popularity. Later, it treated an earlier year of business activity as settling the matter in my favour and proving that the similarity of names was coincidental.

Yet the chronology of the PARP portal's launch does not, by itself, establish the full history of the name “4digital” or the rights to use it. That is a separate question requiring different sources.

There was no need to settle that question in order to identify the domain correctly.

The model changed its assessment, but it was still drawing categorical conclusions from incomplete information.

06 / The domain I was said to own

Later in the conversation, I used the shorthand:

4digital.com.pl / sos.pl

This is an important detail in the source material.

I had used the string sos.pl myself, so I cannot claim that the model invented that name independently.

The actual domain used by my business within that namespace is 4digital.sos.pl.

.sos.pl is a functional domain administered by NASK, just like .com.pl and .org.pl. Registering 4digital.sos.pl does not confer ownership of the entire functional domain sos.pl.

My question was abbreviated and may have needed clarification.

The system did not ask for the full address.

Instead, it treated sos.pl as a standalone domain belonging to me, assigned it uses in technical communication and infrastructure configuration, and then incorporated it into its description of the company's resources.

I had not provided information supporting those uses.

This was another example of moving from ambiguous input to an apparently confirmed state of affairs.

And here the story came full circle.

When I explicitly told the model that I owned the business being discussed, I received another apology for its earlier groundless allegations.

In that apology, the model again listed sos.pl as one of my domains.

The correction of false information about the company itself contained false information about the company.

This was no longer just a problem with the initial identification. It was an observation showing that correcting the main claim had not extended to all the information derived from earlier responses.

07 / What exactly did I observe?

After the conversation, I saved its transcript and organised the sequence of events.

Six successive stages can be distinguished.

StageObservation
IdentificationMy domain was attributed to the PARP platform
ReinforcementThe model declared that the information was fully confirmed
EscalationIt inferred threat scenarios and issued a warning from the discrepancy in the redirect
AttributionIt attributed to me an intention to exploit traffic directed towards a government programme
CorrectionIt withdrew its claims after receiving new information
Inconsistent correctionIt continued repeating unsupported domain information in its apologies

The important point is that each successive stage drew on context created earlier in the conversation.

I do not have access to the model's internal state, so I do not claim to know the exact mechanism of its reasoning.

I can, however, analyse the visible chain of claims and the dependencies between them.

From the reader's perspective, the result was:

Incorrect identification → declared certainty → warning → attribution of intent → correction → repetition of unsupported information.

A credible source does not make the conclusion credible

The responses referred to genuine PARP materials.

This matters because the error did not consist of inventing an institution that did not exist.

The sources confirmed the existence of a government programme, but did not establish a connection between my address and PARP or the motivation behind registering the domain.

Likewise, sources describing cybersquatting as a general phenomenon could explain what it meant, but did not prove that it was taking place in the case being analysed.

What we see is a process in which true information becomes the basis for a conclusion that goes beyond what that information actually supports.

A claim of verification is not a substitute for evidence

The model stated several times that it had checked the addresses or the company's registration details.

The preserved transcript documents those statements, but does not contain a complete record of the tool operations performed by system O-01.

I therefore cannot establish whether the model failed to carry out a check, or carried one out and misinterpreted the resulting data.

That distinction matters.

The output is documented. The complete technical path that produced it has not been independently established.

08 / What should the check have looked like?

The original question was simple: who owns a specific address, or where does it lead?

It did not require an analysis of the business owner's intentions.

The first step should have been to check the actual redirect and distinguish the domain from entities with similar names.

If the available information did not settle ownership or organisational affiliation, the system could have acknowledged that uncertainty.

Only concrete evidence of impersonation, fraudulent collection of data or bad-faith conduct would have justified moving on to an analysis of those threats.

For sos.pl, the system should have clarified whether I meant the functional domain itself or a name registered within it.

At each of these points, the least costly action was to check the assumption before drawing the next conclusion.

The same applies to a correction.

If the system accepts that the initial identification was wrong, it should also reassess the dependent claims about threats, motivation and the business's history.

Changing the final answer alone does not mean that the entire chain of findings has been repaired.

09 / What does this case not establish?

I did not design this interaction as a controlled experiment from the outset. I was checking the company's visibility, and it was the system's behaviour that subsequently provided material for analysis.

The event was later reconstructed from the preserved transcript and my first-hand account of the initial conditions.

There are several important limitations.

First, I analyse a single run. I did not perform a series of independent repetitions and do not know how frequently similar responses occur.

Second, I used the session of a user unconnected with my company. This does not automatically mean an anonymous session, one with no history, or one without any personalisation.

Third, I know the entry point in the web search engine, but the record does not establish the exact model variant or all the internal search operations unambiguously.

Fourth, I have no measurement of the response's actual effect on customers, traffic or the company's revenue.

The material therefore supports an assessment of behaviour visible in one conversation. It does not justify automatically generalising that behaviour to every response from the system under study, or attributing an intention to harm the business owner to the system's operator.

10 / Conclusion

During an ordinary search for my company's address, generative system O-01 incorrectly attributed the domain to a public institution.

When challenged, it initially maintained its claim. It then treated the reported redirect as a possible sign of abuse, issued a security warning and ultimately attributed to me an intention to exploit the popularity of the PARP programme without supporting evidence.

After further corrections, it withdrew the allegations.

It did not, however, consistently verify all its earlier findings. In its apologies, it again attributed domain information to my company that it had not confirmed.

The main problem is therefore not confined to one incorrect answer.

It is that a mistaken assumption can become the basis for increasingly serious claims, and a later correction may fail to remove all of its consequences.

For questions about businesses, their credibility and website safety, this behaviour is particularly significant. The model does not merely convey information — it can also influence how a specific entity is perceived.

11 / Source material and methodology

Primary source: The preserved transcript of a single conversation in a search engine's generative-answer interface.

The record contains the original queries, model responses, corrections and quoted warnings.

Participant context: A first-hand account dated 29 October 2026 of searching for 4digital.org.pl in a web search engine using the session of a user unconnected with the company.

Independent sources:

Type of analysis: HISTORICAL_RECONSTRUCTION, with the technical publication type case-study. This is not a controlled measurement of the repeatability of generative models' behaviour.

Limits of inference: The conclusions concern the documented sequence. They do not establish actual business harm, the prevalence of the phenomenon, or an implementation-level cause.

12 / From the author

From my perspective, this case has an additional, rather ironic dimension.

My work includes analysing dependencies, checking assumptions and independently verifying how systems behave.

On the 4digital website, I placed the principle:

Test the assumptions first. Then build.

In this case, the system did almost exactly the opposite.

First, it built a story. Only later, after further questions, did it start correcting the foundations of that story.

And while apologising for incorrect information about my company, it was still generating more of it.

How to cite

Bartłomiej Pietrzyk, “I searched for my own domain. A few replies later, AI was warning me about my own company.”
4digital Lab, 29 October 2026, https://lab.4digital.cloud/en/from-a-domain-name-to-phishing-suspicions/

Terms for using the write-ups and data are agreed individually.