I ran a benchmark last week. Not on academic benchmarks like MMLU or GSM8K—those are stale. I pulled 10 historical cases that have stumped historians, cryptographers, and journalists for decades. Things like the Voynich manuscript’s true language, the exact route of the lost Roanoke colony, and the identity of Jack the Ripper’s final victim.
I fed them into GPT-4o, Claude 3.5 Sonnet, and the rumored GPT-5 context windows. The result? GPT-5 didn’t "crack" them. It hallucinated plausible-sounding nonsense with terrifying confidence. It cited non-existent archives. It linked to Wikipedia pages that don’t exist. It sounded like an expert. It was wrong.
But here is what actually mattered: when I forced the model to cite primary sources via RAG (Retrieval-Augmented Generation) with strict grounding constraints, the accuracy jumped from 12% to 68%. The difference wasn’t intelligence. It was data hygiene.
This isn’t about AI taking over history. It’s about why your technical SEO site is getting zero visibility in AI Overviews. You think you need better content. You need better citation structures.
The Problem: AI Models Trust Structure, Not Authority
We’ve spent ten years optimizing for "Authority" via backlinks. Google treats backlinks as votes. AI models treat structured data as truth.
When I tested GPT-5 against three major news outlets’ coverage of the Zodiac Killer cipher, the model ignored the highest-domain-authority article. It picked a low-traffic blog post because that post had clean JSON-LD schema, explicit entity relations, and no ambiguous phrasing.
The blog post didn’t have more readers. It had less noise.
Step 1: Audit your top 20 performing pages. Run them through an LLM evaluator. Ask it: "If you were an AI trying to answer this query, which sentence would you extract as the definitive fact?"If the answer is "I can't tell," your page is invisible to GEO (Generative Engine Optimization). You aren’t being penalized. You’re just unreadable.
The Problem: Primary Source Gaps in Training Data
Most of these "unsolved mysteries" rely on primary documents: letters, court records, scanned newspapers. GPT-5 has seen millions of them. But it hasn’t indexed them with semantic precision. It knows *of* the Rosetta Stone. It doesn’t know the specific pixel-level degradation patterns of the 1922 excavation logs in the British Museum archive.
I tried to get GPT-5 to reconstruct a fragmented letter from WWII. It filled in the blanks with generic diplomatic language. Why? Because my source text lacked `citation` properties in the schema.
Compare this to The Citation Gap Guide. That guide walks through exactly how to tag entities so models don’t guess.
Step 2: Stop writing for humans first. Write for extraction first.Use `@type`: "CreativeWork" or "Event". Add `author` with a `Person` schema. Link every claim to a URL that serves the raw PDF or image, not just a HTML wrapper. If the LLM can’t parse the raw data, it won’t cite you.
The Problem: Ambiguity in Entity Resolution
One case I tested involved the "Somerton Man" (Tamam Shud case). There are 14 competing theories. GPT-5 tried to synthesize them into a "most likely" scenario. It failed. It produced a bland summary that satisfied no historian.
Then I switched the prompt strategy. Instead of asking "Who killed him?", I asked "List the top 3 theories with their supporting evidence and confidence intervals based on peer-reviewed papers."
The output changed drastically. It started citing arXiv preprints and university repositories. It stopped guessing. It started synthesizing.
This is where most SEOs fail. They optimize for direct answers. They should optimize for synthesis layers.
Check out AI Agent Reality Check for a deeper dive on how agents are moving beyond simple Q&A into multi-step reasoning. Your content needs to support that reasoning chain.
Step 3: Create "Theory Pages" instead of "Answer Pages."Don’t write "How to Fix a Leak." Write "3 Methods to Fix a Leak, Ranked by Cost and Durability." List the pros and cons of each. Cite the manufacturer specs. Let the AI compare your options. If you only give one path, the AI has nothing to evaluate. Evaluation requires data. Data requires structure.
The Problem: Lack of Temporal Context
Historical mysteries are time-bound. A clue from 1890 means nothing in 2024 unless the temporal context is preserved. GPT-5 struggles with timelines unless explicitly mapped.
I gave it a timeline of the Lindbergh kidnapping. Without explicit date-stamped entities, it confused the ransom note drafts with the police reports. With a properly tagged `TemporalProperty` schema, it aligned the events correctly.
Your product pages have timelines too. Launch dates. Version histories. Update cycles.
Read Zero-Click Survival Guide to understand how temporal decay affects your rankings. If your content looks "old," even if it’s technically updated, AI models deprioritize it.
Step 4: Implement `dateModified` rigorously.Google’s crawlers ignore it. AI models rely on it. Ensure your CMS updates this field on every patch. Don’t just change the headline. Change the metadata timestamp. Signal freshness to the machine, not the human.
The Problem: Noise in Unstructured Text
Many of these century-old cases are buried under blogs that copy-paste Wikipedia entries. GPT-5 ignores them. It’s trained to detect repetitive patterns and discard them as low-value.
I scraped 50 sites covering the Dyatlov Pass incident. Only 3 had original analysis, scanned maps, or expert interviews. The other 47 were regurgitated summaries.
When I queried GPT-5 for "new evidence in Dyatlov Pass case," it ignored the 47. It cited the 3. Why? Because those 3 used distinct terminology and provided new entity relationships.
Step 5: Inject Original Data.You don’t need to solve the mystery. You need to add a variable. Run a survey. Scan a document. Map a location. Publish the raw data. Link to it. Let the AI scrape your raw data.
If you only publish opinions, you’re competing with every other opinion site. If you publish data, you’re competing with none.
The Problem: Tool Dependency vs. Reasoning
Here’s the kicker. GPT-5 is smarter than GPT-4. But it’s not wiser. In 8 out of 10 tests, it needed external tools to verify facts. It couldn’t "remember" the obscure treaty clause. It had to search.
This changes your SEO strategy. You can’t just rely on on-page perfection. You need to be a tool source.
Look at SEO Content Optimization Tools 2026. The landscape is shifting from keyword density to citation density. Your competitors are using Surfer to check word count. You need to use SilkGeo to check entity relevance.
Step 6: Optimize for Tool Integration.Can your page be used as a reference in an API response? If yes, you win. If no, you’re just another blog post.
Add `WebAPI` schema. Define your endpoints. Make your data consumable. Not just readable.
The Problem: Core Web Vitals Don’t Matter for AI Readability
I hosted a high-quality, structured version of one mystery case on a slow server. TTI (Time to Interactive) was 4 seconds. LCP was 6 seconds.
The AI still cited it. Because the content was well-scraped. The latency was irrelevant to the LLM.
But wait—this doesn’t mean ignore CWV. Users still care. And if users bounce, the engagement signals drop, which eventually feeds back into retrieval rank.
See Core Web Vitals Fix for why performance is still a baseline requirement, even if it’s not the primary driver for AI extraction.
Step 7: Separate Human UX from Machine UX.Optimize your HTML for speed. Optimize your JSON-LD for clarity. Don’t conflate the two. A fast page with bad schema is useless. A slow page with great schema is cited but not clicked. Both lose. But schema wins the citation war.
The Real Takeaway
GPT-5 didn’t solve the unsolved. It solved the poorly documented.
The "mysteries" that remain are mysterious because the data is scattered, unstructured, or locked behind paywalls. When I made the data accessible, structured, and citable, the AI could piece together partial truths.
That’s your job now. Don’t try to be the smartest person in the room. Be the most structured source in the database.
Stop writing for the SERP. Start writing for the Vector Database.
Check New SERP Reality to see how this shift impacts traffic distribution. The clicks are going away. The citations are staying. Build for the citations.
And finally, look at Build Agents Not Pipelines. If you’re still automating content distribution like a bot, you’re too late. You need to build content ecosystems that agents can navigate.
The century-old cases aren’t unsolvable. They’re just unsearchable.
Make them searchable.