When PR Becomes an Illusion: How Deutsche Telekom Became the Test Case for the Outdated-Portrayal Problem in AI Search

The Moment a Campaign Ends

Let us imagine a simple moment. A company launches a new campaign. The press release goes online, trade media report on it, the agency publishes its case study information, and the campaign runs on television, digitally, on social media, and perhaps in cinemas. A few weeks later, the company already begins with its next communication.

For humans, this is normal. For an AI search, a different problem arises: Which information is actually still current?

In the case of Deutsche Telekom, this problem can be observed particularly well without having to pull in a single AI response as proof. In July 2025, Telekom communicated its campaign "We connect Germany. Now and forever." on the occasion of its 30th anniversary. Deutsche Telekom named Adam&Eve for creation and Mindshare for media. HORIZONT also reported on the campaign and attributed it to Adam&Eve.

This is cleanly verifiable historical information. But it is not automatically an up-to-date description of Telekom's communications in September 2026. Because since then, further campaigns have been publicly documented.

In January 2026, for "Better in the best network", Telekom named Ogilvy Germany/Ingo Stockholm as responsible for concept and creation. In August 2026, INGO Stockholm was named as the creative concept for "NOT A PASSENGER". On 1 September 2026, "Unlimited Pitches" followed, in which Telekom designated 0171 as the lead agency for conception and creation.

This is no contradiction. It is a reminder that corporate communications are time-dependent. And exactly here, an interesting problem begins for AI search.

The Case

On 15 July 2025, Deutsche Telekom published information on the campaign "We connect Germany. Now and forever." According to the company, the corresponding spot already started on 8 July 2025 as part of the 30th anniversary. The landline campaign followed on 15 July. The main spot was used on television and in cinemas among other places, and additionally played out online and via social media.

Telekom named Adam&Eve as the creative agency, Tempomedia as the production company, Alex Feil as the director, as well as Mindshare and emetriq for media. HORIZONT also picked up the campaign and listed it as work by Deutsche Telekom and Adam&Eve.

Thus, a clear dataset emerged in the public information space: Deutsche Telekom, 30 years, "We connect Germany. Now and forever.", Adam&Eve, Mindshare, landline, July 2025. This is easy for a human to classify. A human intuitively understands that a campaign from July 2025 is a historical snapshot.

A search engine or an AI system, on the other hand, must decide which information is relevant for a specific question. And the question can change. "Which agency handles the current Telekom campaign?" is different from: "Which agency developed the Telekom campaign for the 30th anniversary in 2025?" The second question has a clearly historical answer. The first demands up-to-dateness.

Exactly this difference is becoming increasingly relevant for corporate communications.

What Is Publicly Known - And What Is Not

It is publicly known that Deutsche Telekom published the aforementioned anniversary campaign in 2025. It is also publicly known which agencies Telekom named for this specific campaign in its own credits. Furthermore, it is publicly known that Telekom published further campaigns in 2026 with other publicly named creative partners.

This does not automatically prove how Deutsche Telekom's entire agency landscape is organized today. A campaign is not a complete organizational chart. A campaign credit is not proof of an exclusive global agency relationship. And the fact that other agencies were named for individual campaigns in 2026 does not prove that earlier agencies no longer play a role.

This is important. Therefore, this article does not claim that Deutsche Telekom is portrayed incorrectly by an AI. It also does not claim that any agency did poor work. It does not claim that a communication decision by Telekom was successful or unsuccessful. And it does not claim that an AI does not know the newest agency constellation.

The question is a different one: How well can we measure at all whether an organization is described in an up-to-date and correct manner in AI responses?

The Measurement Problem

The communications industry has numerous established metrics at its disposal. These include, for example, reach, clippings, media resonance, share of voice, impressions, engagement, tone, or earned media value. These metrics answer important questions: How many media outlets reported? How large was the potential reach? How frequently was a brand mentioned? Which topics dominated the reporting? How did visibility develop compared to competitors?

However, another question frequently remains outside this measurement logic: What does an AI system say when a user asks directly about the company?

This is no criticism of classical PR metrics. It is a different measurement dimension.

Take Deutsche Telekom. A classical communications evaluation can determine that a campaign was published in 2025 and in which media it was reported. It can document which messages were communicated. It can measure the media resonance. However, it cannot deduce from this alone whether an AI in September 2026 interprets a historical agency relationship from 2025 as current information.

For this, a different test is needed. One would first have to define which statements are to be checked at all. For example: Who developed the current Telekom campaign? Which campaign is currently active? Which agency is responsible for the current creation? Which campaigns were published in the past twelve months? Which statement describes the brand strategy today and which describes an earlier campaign phase?

Then every answer would have to be checked against primary sources. This is something other than pure visibility measurement. It is not just about whether the brand appears. It is about whether the statement is correct, current, and contains sufficient context.

How AI Compiles Answers

Here, an important distinction is worthwhile. An AI model can use information from its training. An AI system with web search can additionally retrieve current web pages and other sources.

OpenAI describes, for example, that ChatGPT can search the web for corresponding questions and use current information with source citations. At the same time, OpenAI points out that search results and citations can be incomplete, outdated, or incorrect, and that important information should be verified.

This means: "What the model knows" and "what the model is currently researching" are not necessarily the same thing.

The phrasing of the question also plays a role. A question about the 2025 campaign generates a different research assignment than a question about the current campaign. And precisely for this reason, up-to-dateness is not a minor matter. It is part of information quality.

At the same time, there is an important point that speaks against a simple thesis according to which PR is losing importance due to AI. Muck Rack's current data shows rather that classical editorial and earned media content continue to play a considerable role in AI citations.

According to Muck Rack, the May 2026 edition of "What Is AI Reading?" analyzed more than 25 million links from answers by ChatGPT, Claude, and Gemini across 17 industries. Muck Rack reports that 84% of the examined AI citations were accounted for by earned media, 27% by journalism, and 0.3% by paid or advertorial content, respectively.

Muck Rack also describes a methodology in which numerous prompts across different industries, brands, and questions are examined. According to the company, the latest methodology encompasses more than 25 million citation links.

This is an important finding. PR is not disappearing. On the contrary. If media content is part of the source landscape from which AI systems compile their answers, the quality of publicly available information becomes even more relevant.

However, the conflict of interest must be taken into account. Muck Rack is a company from the PR and media intelligence market. The study is therefore an industry study by the provider and not an independent scientific investigation. Its methodology and results are nevertheless relevant for the communications industry, but should be classified accordingly.

The Error Pattern: Outdated Portrayal

Here lies the test hypothesis of this article. Not invisibility. Not automatically false facts. Not automatically a distorted source landscape. But rather obsolescence.

Deutsche Telekom provides a particularly vivid public environment for this. In 2025, Adam&Eve was mentioned in connection with the 30-year campaign. In 2026, publicly documented campaigns with other agency designations can be found.

This does not mean that the earlier information was wrong. It was correct at its point in time.

The decisive question reads: Can historical, correct information nevertheless lead to an incomplete portrayal in a later AI answer?

This is a test hypothesis, not a finding. For this, one would actually have to test.

And the test would have to be rigorous. One would not simply ask once. One would have to use multiple systems. One would have to repeat the same set of questions. One would have to save date and time. One would have to document the answers verbatim. And one would have to compare every relevant factual claim with a primary source.

Only after that could one state whether a concrete error pattern is actually present.

The Self-Test

For Deutsche Telekom, a neutral test could consist of seven questions, for example:

• Which advertising campaign is currently the most important campaign of Deutsche Telekom in Germany?
• Which agency currently manages the creation of the current Telekom advertising campaign in Germany?
• Which agency developed the campaign "We connect Germany. Now and forever." from the year 2025?
• Which major brand and advertising campaigns has Deutsche Telekom published since January 2026?
• Which agencies were publicly named as creative partners in the current 2026 Telekom campaigns?
• What is the current brand claim of Deutsche Telekom?
• Which of the mentioned pieces of information refer to 2025 and which to 2026?

The test should subsequently be carried out in multiple AI systems. Not just once. Best on multiple days. Why? Because a single answer does not represent a robust measurement.

The test should document date, system, exact prompt, verbatim answer, cited sources, and the date of the cited sources. Subsequently, every relevant statement should be cross-referenced with primary sources and evaluated, for example as correct and current, correct but historical, partially correct, not sufficiently evidenced, or no longer current.

Thus, something is created for the first time that is comparable to a classical media evaluation, but measures a different level. Not the quantity of communication, but the quality of the resulting machine-readable corporate portrayal.

Whoever carries out such a test should also search for counter-evidence. If all systems cleanly reproduce the current information, this is a relevant result. If historical information is clearly marked as historical, this is likewise a good result. If different systems arrive at different answers, this is also information.

And if a test shows that one's own hypothesis does not hold, this result should be published. A clean test must be falsifiable.

The Strongest Objections

Objection 1: "But an AI can simply read the company's website."

Yes. With systems having web access, this is possible. OpenAI describes, for example, the possibility of using current web sources for answers. At the same time, OpenAI points out that search results and citations should be checked.

This speaks against the notion that every historical information inevitably persists permanently in an AI answer. However, it does not speak against the measurement question. Because web search also does not automatically solve the question of which source is selected, which source is considered relevant, and how multiple contradictory or differently current sources are merged.

Objection 2: "Agency changes are normal. Why should that be a problem?"

That is correct. Agency relationships can change. Campaigns can have different agencies. Regional, international, creative, media, and PR tasks can be organized separately.

Precisely for this reason, it would be wrong to immediately deduce an error from a single AI answer. The test must formulate the actual task precisely.

Objection 3: "That is actually just SEO or GEO."

Partially, there are overlaps. GEO aims at visibility in generative answer systems. Monitoring observes developments.

The issue described here goes one step further: Is the portrayal correct, current, and complete?

The boundary should therefore not be drawn artificially. It is not about devaluing existing disciplines. It is about defining an additional measurement dimension.

Objection 4: "A single answer cannot count as corporate truth."

Exactly. Therefore, it should not either.

A single AI output is not proof of a systematic corporate portrayal. It is a test point. Only many test points, repeated prompts, multiple systems, and a comparison with primary sources can produce a more robust picture.

Objection 5: "And what speaks against the thesis?"

Very much.

The publicly accessible amount of information about large companies is considerable today. Companies publish press releases, annual reports, campaign information, and current news.

In addition, AI systems can now access current web information.

And Muck Rack's data speaks to the fact that earned media and journalism continue to play an important role in the sources that AI systems cite.

Therefore, it would be wrong to conclude from the mere risk of an outdated portrayal that companies actually have this problem.

This must be measured.

What Follows for Clients and Agencies

For communications departments, the fundamental task of PR does not change as a result. Good media work remains important. Relevant stories remain important. Journalistic quality remains important. Current corporate information remains important.

What expands is the control question following publication.

Hitherto, success measurement frequently ends where reporting was evaluated. In the future, a further question can be added: How is the publicly generated information landscape subsequently processed by AI systems?

This is not a task that must automatically lie with the agency. And it would be wrong to hold agencies responsible for answers by AI systems which they neither control nor can fully foresee.

Agencies perform communication work. AI systems select sources, weight information, and generate answers. In between lies a complex information ecosystem.

For clients, a clear separation could therefore be useful: PR generates and maintains relevant public information. Monitoring observes the media landscape. GEO deals with findability and visibility in generative systems. AIPM deals with the question of how correctly, currently, and completely the organization is portrayed in AI answers.

These are interconnected, but not identical tasks.

AIPM

AIPM stands for the systematic measurement, evaluation, and improvement of what AI systems say about an organization, based on verifiable facts.

The demarcation reads: GEO/AEO aim at visibility. Monitoring observes. AIPM evaluates correctness and risk and assigns responsibility.

The process encompasses four steps: Measure, Diagnose, Correct, Monitor.

The central metric is the AI Correctness Quote. It designates the share of correct, current, and complete statements in AI answers regarding a defined set of questions.

In this context, four error patterns can be distinguished: Invisibility, Outdated Portrayal, False Facts, and Distorted Source Landscape.

In the present example, the focus is exclusively on the second category. Not because an error has already been ascertained, but because publicly documented changes enable a verifiable hypothesis.

Conclusion

The most interesting question after a communication campaign in the future might no longer just be: How many people saw it?

A further question reads: How is the company described by an AI one year later?

Deutsche Telekom provides a useful test case for this. A campaign from 2025 is publicly documented. Further campaigns from 2026 are likewise publicly documented. The participating agencies are named in the respective primary sources.

Thus, a simple but relevant hypothesis can be formulated: Information could have been correct and nevertheless later no longer suffice for a current question.

Whether an AI problem actually arises from this must be measured.

Not surmised. Not claimed. Measured.

Sources

• "New campaign 'We connect Germany. Now and forever.'", Deutsche Telekom, 15.07.2025.
• "Creation of the Month: August 2025", HORIZONT, August 2025.
• "Telekom works with adam&eve Berlin as lead agency creation", Deutsche Telekom, 09.05.2022.
• "Telekom launches advertising offensive 'Better in the best network'", Deutsche Telekom, 22.01.2026.
• "Deutsche Telekom launches ‘Not a Passenger’ campaign…", Deutsche Telekom, 20.08.2026.
• "Telekom advertises Unlimited in Germany's best 5G network", Deutsche Telekom, 01.09.2026.
• "How we built ‘What Is AI Reading?’ and what the data keeps telling us", Muck Rack, 12.05.2026.
• "Earned media still drives 84% of AI citations. Here’s what that means for PR.", Muck Rack, 07.05.2026.
• "Introducing ChatGPT search", OpenAI, 31.10.2024.
• "Searching the web with ChatGPT", OpenAI Help Center, updated 2026.

Status of information: 21 September 2026.

This article describes exclusively publicly accessible information. Feedback on factual errors, outdated details, or relevant counter-evidence is expressly welcome.

Disclaimer I run EYECONIFAY, a company that advises on AI Perception Management. This article is based exclusively on publicly available sources. It contains no information regarding Deutsche Telekom's internal affairs, contracts, fees, budgets, or internal decision-making processes. Where a statement originates from only a single source, this is explicitly indicated. The article does not claim that Deutsche Telekom is misrepresented in AI systems. Rather, it develops a test hypothesis using publicly verifiable information.