AGI House · The Living Internet · Oct 10, 2026
Get recommended by the AI agents deciding where buyers go — no traffic required.
A simulator of AI answer engines, a bandit that rewrites pages without inventing anything, and a pre-registered test on a real engine — which we report honestly, including that it failed.
Demo video
The problem
"Which cobot should I buy for welding?" now goes to ChatGPT, Claude or Perplexity. The answer names three brands.
In our probes each engine read pages per question before choosing.
A/B testing and tools like Coframe learn from your own visitors. A cobot startup with a few hundred visits a month can't.
How it works
Ask real engines buyer questions, many times. Record pages read and brands named.
A cheap model answers from the same pages; validated against the real engines.
A Thompson-sampling bandit picks one of 7 edit types; an LLM applies it.
A claim checker rejects anything not supported by the original page.
Pre-registered head-to-head on the real engine, held-out questions.
Step 1 · What the engines recommend today
Share of answers naming the brand in its top 3 · 40 cobot buyer questions · Claude Opus 5.5, OpenAI gpt-5.5, Perplexity
Step 2 · Is the simulator any good?
Our first simulator (told to use "only these results") scored r = −0.05: it trusted vendors' own "best cobot" blogs. Letting it blend sources with its own knowledge — as real engines do — fixed it.
Step 5 · The pre-registered result FAIL
Noise: the same unchanged page, asked twice, scored 50% vs 38% (Franka) and 25% vs 38% (Elite).
So: simulated gains measured with 16 runs per edit sat inside that noise; the optimizer kept lucky edits.
Fix: many more simulator runs per edit, more questions, more final trials.
Per company
Numbers only — the companies' own pages and the rewritten versions are not published.
Keeping it honest
| "But…" | What we did |
|---|---|
| Your simulator isn't reality | The final test uses the real Claude Opus 5.5 |
| Any rewrite would help | Compared against a generic AI rewrite and a length-matched placebo |
| You tuned on the test | 15 held-out questions never used for learning |
| You picked the winners | Hypothesis, companies and analysis committed to git first (f7a0b9f, eee5a57, 0e88e8e, e70b303) |
| It's manipulation | Claim checker; no hidden text; the same page for people and bots |
Where it goes
Weekly share of AI answers per engine; alerts when a model update shifts who wins.
The same simulator for /agent pages, MCP endpoints and personal-agent protocols.
Every connected site's real outcomes recalibrate the simulator for everyone.
Map which sources feed each answer — and where you're missing.
The no-invented-claims checker as a guardrail for self-optimizing sites.
Cobots today; B2B SaaS, dev tools and local services next.