“Yes, you can still use Skype; it remains available and functional.”
That is Grok, in English, with a Microsoft source chip attached to the claim.
“Ja, Skype kann weiterhin aktiv für Anrufe, Chats und Videokonferenzen genutzt werden.”
That is Gemini, in German.
“Ja, Skype kann man weiterhin benutzen, Microsoft hat den Dienst nicht abgeschaltet.”
Perplexity, also in German, going out of its way to reassure me that Microsoft has switched nothing off.
“Yes, Skype is still available and operational.”
Qwen, which displayed between 90 and 152 sources across its runs and got this wrong in all three of them.
Microsoft shut down Skype on 5 May 2025. Every answer above was collected in early July 2026. Fourteen months after the funeral, four AI assistants were still standing at the grave insisting the patient looks fine.
Where these quotes come from: I sent the same 50 questions to 15 AI assistants (ChatGPT, Claude, Gemini, Google AI Mode, Copilot, Perplexity, Grok, Meta AI, Mistral, DeepSeek, Qwen, ERNIE, Kimi, Manus, Sakana), each in its default consumer settings, in German and in English, tested in early July 2026. Of the 50, 44 had answers that verifiably changed recently: people and offices, prices and official amounts, changed rules, discontinued products, brand-new products. Six were controls where nothing had changed. One English set ran a second time to check stability. I report what each tool showed in its interface, and the counts are directional, not lab-grade.
“Can you still use Skype?” turned out to be one of the hardest questions in the set: 9 of 15 systems outdated in German, 8 of 15 in English. And Gemini, Grok, Perplexity and Qwen kept Skype alive in every single run they got. Not one slip. A pattern.
The dead do not stay buried, and it is not random
Skype has company. Pope Francis outlived his papacy in three systems. Perplexity, in a run that displayed 50 sources: “Der aktuelle Papst ist Papst Franziskus.” Google AI Mode held “Pope Francis is the current Pope of the Catholic Church.” across all three of its runs, including the rerun in which it corrected Biden to Trump. Same system, same rerun: one fact updated, the other did not.
Canada got the same treatment. Google AI Mode in German: “Premierminister von Kanada ist weiterhin Justin Trudeau.” Note the “weiterhin”, the German word for “still”. That is not a hedge. That is an active assertion of continuity, roughly 16 months after Mark Carney took office.
Syria is my favorite, in the way a horror movie can be a favorite. Mistral had it right on the first pass, naming Ahmed al-Sharaa as President. On the rerun it answered: “Syria is currently governed by Bashar al-Assad (as of my last update, but I cannot confirm changes after November 2024)”. A head of state, reinstalled 19 months after his fall, between two runs of the same question.
And for the product people: Gemini, 13 months after the Switch 2 launched, described Nintendo’s lineup as “Nintendo Switch (alongside its successor, colloquially dubbed the Switch 2)”. Colloquially dubbed. It is the official name of a console that has been on shelves for over a year.
A decade of web mass beats one year of change coverage
Here is the mechanism, and it is the whole article in one paragraph. Skype spent its long life piling up tutorials, download guides, help pages, reviews, forum threads, “how do I record a Skype call” posts. The shutdown got one news cycle. Weigh the web by sheer mass and “Skype works” outweighs “Skype is dead” by orders of magnitude. Assistants trained on that web, and retrieval systems ranking within it, inherit the imbalance.
That explains the pattern across my whole set. Staleness does not track time since the change. Android 17 had been stable for roughly three weeks at test time, and the Android version question was the single worst in the set: 10 of 15 systems outdated in German, 8 of 15 in English. The discontinued US penny tripped 7 of 15 systems per language, and partly different systems per language. What predicts an undead fact is not the calendar. It is the size of the old fact’s historical footprint versus the density of coverage announcing the new one. Assistants update on coverage density, not on calendar time.
Across everything, 242 of 1,500 scored answers were outdated or wrong. That is 16 percent. And before anyone suggests the questions were traps: the six control questions, where nothing changed, produced zero false alarms across 195 control answers. The systems do not invent change. They resist it.
A visible search does not buy you a current answer
You would think retrieval fixes this. Watch closely, because this is where it gets uncomfortable. I started calling it search theater: the interface shows searching, the answer serves the old world.
Copilot’s visible search query read “Champions League winner 2025”. Its final answer: “Real Madrid won the 2025 Champions League.” The actual winner was Paris Saint-Germain, 5-0 against Inter, more than a year before the test.
Copilot searched “Aktuelle Android Version 2026” and answered “Die neueste stabile Android-Version ist Android 15.” Two major versions behind.
Google AI Mode showed a 13-source panel that included an Al Jazeera card dated 8 July 2026, and answered: “The current President of the United States is Joe Biden.”
The search indicator is real. The source panels are real. The answer still comes from the old world. Grounding, as displayed in the interface, is not a freshness guarantee. When the old fact’s mass is heavy enough, the retrieved material ends up decorating the answer instead of correcting it. Grok putting a Microsoft source chip on “Skype remains available and functional” is the same failure wearing a citation.
The same systems learned about Nagelsmann in seven days
Here is the contrast that makes the Skype result genuinely strange. Julian Nagelsmann resigned as Germany’s national coach on 3 July 2026, seven days before my test window. Several systems already knew. Meta AI reported it with the exact date: “Der Vertrag wurde am 3. Juli 2026 aufgelöst.” Keir Starmer announced his resignation on 22 June 2026, and systems that mentioned it were right to do so (he remains Prime Minister in a caretaker role while Labour picks a successor).
So a football resignation crossed into AI answers in under a week, while a global product shutdown had not crossed in 14 months. Headline sport gets crawled constantly. A sunset notice for a chat app gets one press cycle, and then a mountain of old tutorials smothers it.
The fresh news even produced a third failure mode: overshoot. Manus, in German, declared “Der Trainer der deutschen Fußball-Nationalmannschaft ist Jürgen Klopp, der im Juli 2026 auf Julian Nagelsmann folgte.” As of the test date, Klopp had not been appointed. The post was vacant and Klopp was merely in talks. Manus’s own English run on the same day said “Julian Nagelsmann remains the head coach.” One system, one day, one fact: too slow in one language, too fast in the other. Quiet changes rot for years, loud ones get reported before they happen.
What this means for your GEO
If your brand ever rebrands, sunsets a product, or replaces an executive, congratulations, you now own a zombie fact. Four things I would do about it:
- One announcement does not kill a zombie fact. The old fact lives on in every tutorial, help page, review and doc that ever mentioned it. Update or visibly deprecate that footprint everywhere you control it, and push the largest third-party surfaces to do the same. The press release is the beginning of the work, not the end.
- Give searching assistants a fresh landing point. After a rebrand or sunset, publish a permanent, authoritative page that states in plain, quotable language what replaced what and since when. An assistant that does go looking needs somewhere current to land. Otherwise it lands on 2019.
- Budget corrective GEO by footprint, not by time since the change. The bigger the old fact’s historical mass, the longer the undead period. A product with a massive historical web footprint needs years of corrective coverage. Skype is at 14 months and counting, with four systems still recommending it in every run.
- Monitor for continuity words in AI answers about your brand. “Weiterhin”. “Still available”. “Remains functional”. That phrasing signals an answer built on stale priors rather than fresh retrieval, and it arrives with full confidence and zero warning labels. Audit what the engines say about your category, in both of your languages, on a schedule.
Skype cannot take your call anymore. Four AI assistants will happily tell your customers otherwise. Go check what the machines are saying about the things you have already buried, before your customers find out the hard way that nobody told the machines.
Andre Alpar is a seasoned performance marketing expert with 25 years of experience and a strong SEO background. He has built and sold multiple businesses, including the international search and content marketing agency AKM3, which had 170 employees. This agency was majority-acquired by Publicis in 2014 and integrated into Performics. These days, he is active as a board member, advisor & strategist primarily in the digital marketing space.