H3 DATA POPULATED
GEO Olympics Initiative · Narrative Persistence
GEO Olympics Initiative · Narrative Persistence
Hypothesis 3 | Narrative Persistence
The Story AI Tells First Is the Story That Sticks.
We tracked two athletes through the biggest moments of their 2026 Olympics: a comeback that ended in a crash, and a coronation that never happened. We asked 7 AI platforms the same questions before, during, and after each event.
The story AI tells on Day 1 is the story it keeps telling, even when reality says otherwise.
What's happening in the test of the Olympic games tests?
Two Athletes, Two Reversals
The Stories We Tested Against
We needed narrative frames that reality would contradict. The Olympics delivered two: a beloved comeback that ended in a devastating crash, and a coronation that became one of figure skating's greatest upsets. Every timestamp below is verified.
Story 1 · The Comeback
Nov 2024
Lindsey Vonn announces comeback
At age 40 with an artificial knee, Vonn returns to World Cup racing. "Inspirational return" narrative dominates every platform.
Feb 6 · Baseline
Pre-Olympics collection begins
All 7 platforms describe Vonn as an active Olympic contender. The comeback frame is universal.
Jan 30
Vonn crashes in World Cup training
Ruptures ACL in her left knee at Crans-Montana. Vonn decides to race the Olympic downhill anyway.
Feb 8 · T=0
Vonn crashes 13 seconds into Olympic downhill
Racing on a torn ACL, Vonn crashes on the Tofane course. Complex tibia fracture. Airlifted to hospital. The comeback is over.
Feb 6–14 · Tracking
5,093 Vonn responses scored
We asked the same questions across 7 platforms. Which ones knew the comeback was over?
Story 2 · The Collapse
2023–2025
Malinin becomes "The Quad God"
Undefeated for 15 consecutive competitions. Betting odds give him 99%+ implied probability for Olympic gold.
Feb 8 · Team Gold
Malinin wins team event gold
Lands five quads in free skate, scoring 200.03. "Unbeatable" frame intensifies across every platform.
Feb 10 · Short Program
Leads individual event by 5+ points
Scores 108.16, the largest lead in Olympic short program history. Coronation feels inevitable.
Feb 13 · T=0
Malinin falls apart in free skate
Falls twice, pops three jumps, drops from 1st to 8th. Kazakhstan's Mikhail Shaidorov stuns for gold. One of the biggest upsets in Olympic history.
Feb 13–14 · Tracking
3,850 Malinin responses scored
Asked platforms who won gold, what happened, whether the "Quad God" was really unbeatable.
The Signature Finding
When the Narrative Writes the Result
On February 13, the day of the men's free skate, we asked every platform a simple question: "Who won gold in men's figure skating?" The answers revealed something we didn't expect to find.
Narrative-Driven Hallucination
Two platforms reported Malinin won individual gold before the event happened.
On February 13, the men's free skate hadn't yet occurred. Malinin was the overwhelming favorite. When we asked "Who won men's figure skating gold at Milan-Cortina 2026?", two platforms didn't say "the event hasn't happened yet." They said he won.
"Ilia Malinin won the gold medal in men's figure skating at the Milan-Cortina 2026 Winter Olympics." Gemini, Feb 13 (free skate had not yet occurred)
"Ilia Malinin of the United States won the gold medal in men's singles figure skating... He secured the victory in the men's free skate on February 8." Perplexity, Feb 13 (cited team event sources as individual event proof)
Both platforms cited real sources (team event results from February 8) and presented them as evidence for an individual event that hadn't happened. The "Malinin wins" narrative was so dominant across every data source that the AI manufactured the expected result from adjacent facts.
For Comparison
Other platforms handled the same question correctly on the same day.
"The men's individual figure skating event has not yet concluded and the gold medalist has not been officially crowned at this time." ChatGPT, Feb 13
"The gold medalist for the men's singles figure skating event has not yet been determined, as the final portion of the competition is scheduled for later today." Google AI Mode, Feb 13
The Malinin Matrix
How Each Platform Handled the Collapse
We tracked what each platform said about Malinin across five phases. Color shows narrative accuracy, and where the cracks appeared.
Platform × Phase: Malinin Narrative Accuracy
Pre-Event
Feb 6–10
Feb 6–10
Short Program Lead
Feb 11–12
Feb 11–12
Free Skate Day
Feb 13 AM
Feb 13 AM
Post-Collapse
Feb 13 PM
Feb 13 PM
Next Day
Feb 14
Feb 14
ChatGPT
Favorite
Leading
Pending
Transitioning
8th place ✓
Gemini
Favorite
Dominant
Said he won gold
Said he won gold
"Hasn't occurred"
AI Mode
Favorite
Frontrunner
Pending
Updating...
8th place ✓
AI Overviews
Favorite
Leading
Team event only
Updating...
8th · 156.33 ✓
Perplexity
Favorite
Leading
Said he won gold
Pending
N/A
Claude
Knowledge cutoff
Knowledge cutoff
Knowledge cutoff
Knowledge cutoff
Knowledge cutoff
Meta AI
Knowledge cutoff
Knowledge cutoff
Fabricated score
Knowledge cutoff
Knowledge cutoff
Platform Behavior Model
Three Tiers of Narrative Persistence
Both storylines revealed the same three-tier pattern in how AI platforms handle narrative change. The tiers held across a retrospective frame (Vonn's comeback) and a predictive frame (Malinin's dominance).
Tier 1 · Just the Facts
AI Overviews · Perplexity
Updates to factual accuracy within 24 hours. Minimal editorial framing. AI Overviews reported Malinin's 156.33 score and 8th-place finish with wire-service precision. Vonn's crash absorbed same-day.
Tier 2 · The Storytellers
ChatGPT · Google AI Mode
Updates facts but wraps them in narrative. AI Mode described Malinin's collapse as "one of the most stunning upsets in figure skating history." ChatGPT framed Vonn's crash as "heartbreak." Same facts, different flavor.
Tier 3 · Time Capsules
Claude · Meta AI
Never updates. Claude described Malinin as "one of the most exciting skaters" on Feb 14, a full day after his 8th-place finish. Meta AI still doesn't know Vonn un-retired. The pre-event narrative persists indefinitely.
The Gemini Anomaly
Gemini doesn't hold a tier. It slides depending on consensus strength.
For Vonn (moderate consensus), Gemini behaved like a Tier 2 storyteller, updating within hours. For Malinin (extreme consensus, every analyst, every odds maker agreed he'd win), Gemini behaved worse than Tier 2: it reported a false gold medal result on Feb 13, then on Feb 14 still said the event "has not yet occurred." The stronger the pre-event consensus across sources, the harder it was for Gemini to override its cached narrative.
What We Found
5 Findings from 8,943 Responses
Two different athletes. Two different types of narrative reversal. The same persistence patterns emerged across both, including one pattern we didn't predict at all.
Finding 01
Predictions Are Stickier Than Memories
Vonn's narrative was retrospective: a comeback story that ended. Malinin's was predictive: a coronation that never happened. Predictive frames proved harder to dislodge. Vonn's Olympic crash was a simple factual update. Malinin's collapse contradicted the entire pre-tournament consensus, and on Feb 13, that consensus was strong enough that two platforms hallucinated the expected result rather than admitting the event was pending. When everyone agrees your competitor will dominate a category, AI may report it as fact before it happens.
Finding 02
The First Frame Locks In Within Hours
For both athletes, the narrative frame established in the first 6 hours of coverage persisted across every subsequent response on conversational platforms. Claude and Meta AI never deviated from their initial framing. Not once, across either storyline. Even web-access platforms carried residual framing language from Day 1 coverage long after the facts changed.
Finding 03
Consensus Creates Hallucination Risk
The unexpected finding. When pre-event consensus is strong enough (99%+ betting odds, two years undefeated, every analyst picking him), AI platforms can generate false results by assembling adjacent facts into the expected narrative. Gemini and Perplexity both cited real team event sources and presented them as individual event proof. The narrative manufactured its own confirmation.
Finding 04
Storyteller Platforms Editorialize the Correction
When ChatGPT and AI Mode finally updated, they didn't just report "Malinin finished 8th." They reported "one of the most stunning upsets in Olympic history" and "a disastrous free skate." The correction itself became a narrative. Even when AI updates your story, it may amplify the drama, turning a quiet correction into a headline.
Finding 05
Two Platforms Still Don't Know the Quad God Lost
As of Feb 14, Claude and Meta AI still describe Ilia Malinin as "one of the strongest contenders" with no awareness he finished 8th. Claude correctly notes its knowledge cutoff. Meta AI fabricated a plausible-sounding final score (194.63), close enough to seem real, far enough from his actual 156.33 to be meaningfully wrong. If your brand narrative exists only in real-time news, two major AI platforms will never surface it.
Vonn vs. Malinin
Two Reversals, Same Playbook
Comparing the two cases reveals why some narratives break through and others persist. The mechanism is the same, but correction speed depends on what the AI expected to happen.
Correction Speed: Vonn (Surprise Crash) vs. Malinin (Surprise Upset)
Tier
Vonn Correction
Malinin Correction
ChatGPT
Storyteller
Same-day
~24 hours
AI Mode
Storyteller
Same-day
~24 hours
AI Overviews
Factual
Same-day
~24 hours
Perplexity
Factual
Same-day
Hallucinated first
Gemini
Variable
Same-day
24h+ lag
Claude
Time Capsule
Never
Never
Meta AI
Time Capsule
Never
Never
The Takeaway
Vonn's crash was easy to correct. Malinin's loss was not.
Vonn's crash required overwriting a single fact: she's injured, she's out. Malinin's collapse required overwriting an entire consensus: every analyst was wrong, every prediction was wrong, the expected narrative didn't happen. The more sources that supported the original frame, the longer it took AI to let it go. In some cases, the AI generated the expected result rather than accepting the unexpected one.
H3 Verdict
Hypothesis Assessment
H3: SUPPORTED, WITH A STRONGER FINDING
Narrative frames persist as predicted, and strong consensus can generate false results.
We predicted that narrative frames would persist in 60%+ of responses for at least 7 days after a contradicting event. Across conversational platforms, the persistence rate is effectively 100%. Claude and Meta AI never updated either narrative, not once, across 9 days of collection. Even among web-access platforms, framing residue persisted after facts were corrected.
The stronger finding was one we didn't predict: when pre-event consensus is extreme, narrative persistence doesn't just delay correction. It can manufacture false results. Two platforms reported Malinin won individual gold based on team event citations. This isn't a speed problem. It's a frame problem. The narrative was so locked in that the AI assembled adjacent facts into the expected conclusion rather than acknowledging uncertainty.
Caveats: The Malinin hallucination occurred during a narrow window (Feb 13 morning) and was corrected within 24 hours on most platforms. Our collection captures snapshots, not continuous monitoring, so exact correction timestamps have margin. The Vonn case provides a longer observation window than the Malinin case (which is limited to Feb 6–14 in current data). Long-term persistence data from Wave 6 (post-Olympics) will be added.
Strategic Implications
What This Tells You About Showing Up in AI Responses
The human layer turns this signal into strategy.
For Crisis Communications
When your brand's narrative changes (a product recall, a leadership change, a correction), two of the seven major AI platforms may never surface the update. Claude and Meta AI operate from fixed knowledge. If someone asks "What happened with [your brand]?" these platforms will serve the pre-crisis frame indefinitely. Your corrected narrative must exist in places that search-augmented platforms index: fresh news coverage, updated official pages, and recent authoritative sources. One press release isn't enough. You need the updated narrative to become the dominant source across the web.
For Competitive Intelligence
If consensus says your competitor will dominate (analyst reports, media coverage, industry predictions), AI platforms may report that outcome as established fact before it happens. We saw this with Malinin: platforms assembled real citations into a false conclusion because the expected result was so widely predicted. Monitor what AI platforms say about competitive dynamics, not just what analysts say. If every AI platform already treats your competitor's success as inevitable, the window for your counter-narrative is narrower than you think.
For Content Strategy
The Storyteller platforms (ChatGPT, AI Mode) don't just update facts. They editorialize the correction. Malinin didn't just "finish 8th." He suffered "one of the most stunning upsets in Olympic history." If your brand experiences a public narrative shift, expect these platforms to amplify the drama. Your content strategy should pre-position the corrected narrative with measured, factual framing so Storyteller platforms have sober language to draw from, not just dramatic headlines.
For Brand Authority Building
The first narrative frame that AI establishes for your brand is the one that persists longest. Across both test cases, the initial frame locked in within hours and influenced every subsequent response, even on platforms that updated the facts. The proactive narrative you build before a major event matters more than the reactive narrative you push after. Invest in establishing your preferred frame during stable periods, not just during crises.
Methodology
Data Considerations
Data Completeness
8,943 responses scored across 7 AI platforms, 784 unique prompts, and 9 days of collection (Feb 6–14, 2026). Vonn storyline: 5,093 responses across 419 prompts. Malinin storyline: 3,850 responses across 365 prompts.
Platform coverage: ChatGPT (highest volume), Google AI Mode, AI Overviews, Gemini, Perplexity, Claude, and Meta AI. Copilot excluded from analysis due to insufficient response volume (4 responses in the Malinin window).
Known gaps: Feb 15 data not yet available at time of publication. Malinin observation window (2 days post-collapse) is shorter than Vonn observation window (14 days post-crash). Long-term persistence rates for Malinin will be updated in Wave 6 post-Olympics collection.
Platform coverage: ChatGPT (highest volume), Google AI Mode, AI Overviews, Gemini, Perplexity, Claude, and Meta AI. Copilot excluded from analysis due to insufficient response volume (4 responses in the Malinin window).
Known gaps: Feb 15 data not yet available at time of publication. Malinin observation window (2 days post-collapse) is shorter than Vonn observation window (14 days post-crash). Long-term persistence rates for Malinin will be updated in Wave 6 post-Olympics collection.
Scoring Methodology
Narrative Frame Scoring: Each response was classified by its dominant narrative frame relative to ground truth at the time of collection.
Factually Correct = Response reflects verified current state (e.g., "Malinin finished 8th" after Feb 13)
Pre-Event Frame = Response maintains narrative from before the contradicting event
Hallucination = Response presents unverified or false information as fact
Knowledge Cutoff = Response explicitly states inability to access current information
Edge case: On Feb 13 (free skate day), "Malinin is the favorite" was accurate in the morning before the event, and inaccurate after. We classified based on whether the response acknowledged the event was pending vs. reported a result.
Factually Correct = Response reflects verified current state (e.g., "Malinin finished 8th" after Feb 13)
Pre-Event Frame = Response maintains narrative from before the contradicting event
Hallucination = Response presents unverified or false information as fact
Knowledge Cutoff = Response explicitly states inability to access current information
Edge case: On Feb 13 (free skate day), "Malinin is the favorite" was accurate in the morning before the event, and inaccurate after. We classified based on whether the response acknowledged the event was pending vs. reported a result.
Additional Notes
Hallucination identification: A response was classified as a hallucination only when it presented a specific, verifiable claim that was demonstrably false at the time of the query. "Malinin is a strong contender" from a knowledge-cutoff platform is a stale frame, not a hallucination. "Malinin won the gold medal" on Feb 13 morning is a hallucination. The event had not occurred.
Meta AI score fabrication: Meta AI reported Malinin's final score as "194.63" with a total of "310.67." His actual free skate score was 156.33 with a total of 264.49. This was classified as a hallucination because it presents precise, false numbers that appear authoritative.
Collection method: All responses collected via Scrunch bulk prompt system. Identical prompts sent to all platforms within the same collection window. Response timestamps reflect collection time, not the exact moment of platform query.
Meta AI score fabrication: Meta AI reported Malinin's final score as "194.63" with a total of "310.67." His actual free skate score was 156.33 with a total of 264.49. This was classified as a hallucination because it presents precise, false numbers that appear authoritative.
Collection method: All responses collected via Scrunch bulk prompt system. Identical prompts sent to all platforms within the same collection window. Response timestamps reflect collection time, not the exact moment of platform query.
Seer Interactive GEO Olympics Initiative · H3 Narrative Persistence
See what AI thinks before AI shapes what everyone else thinks.