The open experiment
I am running our own producton myself, in public.
RankAgent is OmniforceAI's product for getting a business named in AI answers. The honest way to show it works is to point it at a domain with nothing behind it and publish every result, including the bad ones. That domain is this one.
Ten questions. Three models. Asked fresh each month in a signed-out session, no personalisation, no memory. A square turns green only when the model names joshuavaz.com or quotes an article from it.
- Latest run
- 0 of 30
- checks where a model named this site
- Runs recorded
- 1
- baseline 16 Sept 2026
- Change since baseline
- Baseline
- one run so far, nothing to compare
16 Sept 2026Baseline
0 of 30 cited
| Question | ChatGPT | Claude | Perplexity |
|---|---|---|---|
| App security check | |||
| What is an FDE | |||
| Showing up in ChatGPT | |||
| RAG versus memory | |||
| Vibe-coding mistakes | |||
| Prompt injection tests | |||
| Pre-launch checks | |||
| Where the bill goes | |||
| Vendor questions | |||
| Demo or product |
Named instead of me
- owasp.org ×9
- anthropic.com ×6
- simonwillison.net ×6
- ycombinator.com ×6
- a16z.com ×6
- snyk.io ×3
- veracode.com ×3
- palantir.com ×3
Day one. The domain had one article on it and no inbound links, so every square being red is the expected result, not a disappointment. This is the line everything after it gets measured against. Each question was asked once per model in a signed-out session with memory and custom instructions switched off.
How the check is run
By hand, once a month. The models have no stable public interface for citation checks and their terms differ, so a script that scrapes them would be both fragile and rude. Instead the script prompts me for each answer, I paste what the model actually said, and it writes a dated file into the repository. The raw files are in the open source of this site, so the scoreboard cannot be quietly edited after the fact.