Every website grader on the internet is the same tool wearing a different logo.
Someone writes a checklist — page speed, headings, mobile friendliness, keyword density — and the tool marks your site against it. The checklist is the product. The judgment was never the machine's; it was the author's, encoded in a rubric and shipped with a nice progress bar on top.
I wanted the opposite, so I built 3drake.com.
The question
Not "is your meta description filled in". Something harder, and far more interesting:
When an AI model looks at your website with nobody telling it what to care about, what does it understand? What does it think it knows? And what does it end up believing?
That is the entire product. You give it a URL. It gives you the answer, from several models at once.
How it works
You paste a domain or a URL. That is the whole input. No account, no signup, nothing else to fill in.
Each model goes and fetches your website itself. Not a scrape I hand over — the model uses its own web tooling, live, at that moment. So it is judging what your site is right now, not what it happened to memorise during training. Rewrite your homepage tonight and tomorrow's run can come back genuinely different.
Then it is left alone with it. No criteria, no checklist, no scale, no worked examples. The model decides for itself what matters about this particular website, judges it against that, and writes what it thinks.
And it gives the site a number. The number has no limit in either direction. It can be 0. It can be 1. It can be 1,000,000. That is the model's call, and I do not intervene.
That last detail is the one that took the longest to accept. The instinct is to write "score from 1 to 10" — and the moment you do, every model converges on a comfortable 7 and you have learned nothing at all. Remove the scale and the model has to invent one. The scale it invents turns out to be one of the more revealing things it produces.
Three models, one submission
Every submission runs through Claude, GPT and Gemini at the same moment, on the same question, with deep reasoning enabled. Not the cheapest tier available — the whole idea collapses if the model is pattern matching instead of actually thinking.
You could open three chat windows and paste the link yourself. But that is three separate errands, and you would be comparing three answers gathered at three different moments. Here all three are asked the same thing about the same website in the same minute.
And I never average them. Everyone tells me to. A single clean number would look far more professional.
The disagreement is the finding. When one model says 12 and another says 40,000 about the same page in the same minute, that gap carries information a tidy average would destroy. They frequently disagree not about whether a website is good, but about what the website is for — and reading that is worth more than any score.
Which is why the reviews matter more than the number. AI can be wrong, confidently and in detail. The interesting question was never "did I get a good score". It is: what did the model think my website was? What did it assume I sell? Who did it decide I am for? People keep telling me the review surfaced something about their own positioning that no analytics tool ever has.
The leaderboard is random and I am not hiding it
Scores are summed and the total puts you on a public board. There is no domain authority, no backlink weighting, no traffic signal, no paid placement.
A personal blog can sit above a company worth billions, purely because of what three models happened to believe on the day. That is not a bug I intend to fix — it is the only reason a board like this is worth looking at. Anyone has a real shot at the top, and sites that land there pick up real clicks.
Two rules keep it sane: one URL means one entry on the board, and a score locks for 24 hours before the site can be re-scored. There is no ownership verification, so re-scoring is the only remedy — and because the models genuinely re-fetch, a site that actually improved can genuinely climb.
The honest part
This is stated up front on the site rather than left for you to discover.
3Drake is a subjective AI opinion, and I have built nothing to soften that. A genuinely excellent website can be handed a 0. A rough one can be handed a million. There is no floor, no ceiling, and no rubric to appeal to.
AI can make mistakes. The reviews are opinions, not an audit.
If that sounds like a flaw in the tool, it is not — it is the experiment. A grader that always returns a comfortable 7 out of 10 would be measuring nothing at all.
The part that is not really about websites
Worth saying out loud: the websites are not the subject here. The models are.
A URL is just material, and that material comes from whoever decides to paste one in. What I am actually collecting is this — when nobody hands a model a rubric, what does it decide is worth valuing?
Give a model a rubric and you are not measuring the model. You are measuring your rubric, and the model is a very expensive spreadsheet. Take the rubric away and something else surfaces: taste. Each model has one. Run it across thousands of real websites and the shape of it starts to appear.
Try it
It is free, there is no account, and every review on the board is public.
Then tell me what you got. I am especially collecting the results where a model clearly decided your website was about something it absolutely is not about — those are the best ones.
Top comments (1)
Hello Glad to see you, I am Kane Lim from Hong Kong. I have over 10 years of development experience. I am writing this because your post was interesting.
The strongest idea here is treating model disagreement as the signal instead of forcing consensus. I would push this further by capturing the complete inference surface: retrieved DOM, rendered content, tool calls, model assumptions, confidence, score distribution, and semantic classification.
The really interesting extension is longitudinal analysis. Store each evaluation as an immutable snapshot and calculate semantic drift after deployments. If your homepage changes, you could identify exactly which concepts models reclassified, whether positioning became clearer, and where model disagreement increased.
I would also introduce adversarial evaluation. Deliberately test ambiguous navigation, misleading metadata, sparse landing pages, JavaScript rendered content, and contradictory copy. That would expose where each model's perception pipeline breaks.
This feels less like another website grader and more like observability for machine perception. The dataset you build could become extremely valuable for understanding how different models interpret the same web surface.
I would be very interested in exchanging ideas around the evaluation architecture.