What AI thinks of your website: we asked three LLMs, with no rubric at all

Why every website grader is really someone's checklist

Run your site through any of the free graders on the internet and you get the same experience: a score out of 100, a progress bar, and a list of things to fix. Page speed. Heading structure. Alt text. Keyword density.

Underneath, they are all the same tool. Somebody wrote a checklist, and the software marks your site against it. The checklist is the product. The judgment was never the machine's — it belonged to whoever wrote the rubric, and any competent team can rebuild it in a weekend.

That was fine when the only reader that mattered was a search crawler. It is not fine now. A growing share of the people deciding whether to trust your business never see your homepage at all. They ask a model, and they read what the model says back. Which raises a question no rubric-based grader can answer: what does the model actually think your website is?

So we built the opposite of a checklist and put it at 3drake.com.

What 3Drake actually does

You paste a domain. That is the entire input — no account, no signup, nothing else to fill in.

Three frontier models then receive one thing each: your URL. Not a scrape we prepared, not an extract we chose. Just the address. Each model goes and looks your website up for itself, decides on its own what is worth paying attention to, writes what it thinks, and gives the site a number.

Three details are deliberate, and they are what make the output worth reading.

No criteria are supplied. No checklist, no guidelines, no worked examples. The model has to work out for itself what this particular website is for and judge it on that basis.

The score has no scale. It can be 0. It can be 1. It can be 1,000,000. This one took us the longest to accept. The instinct is to write score this from 1 to 10 — and the moment you do, every model settles on a comfortable 7 and you have learned nothing. Take the scale away and the model has to invent one, and the scale it invents turns out to be one of the more revealing things it produces.

The reviews are the product; the score is the headline. The number gets attention. The paragraph underneath it is where the actual information is.

How to see what an LLM says about your website

You can do a rough version of this yourself in about ten minutes, and we would rather you understood the method than treated our tool as a black box:

1. Open ChatGPT, Claude and Gemini in three tabs, each with web access enabled.
2. Give each one nothing but your URL and an open question — what do you understand about this website, what do you know about it, and what do you believe about it?
3. Resist the urge to explain your business. The moment you brief the model, you are reading your own briefing back.
4. Compare the three answers side by side and note where they disagree about what you do, not about whether you are good.

The reason we built a tool for it is that the manual version has two flaws. It is three separate errands, so you are comparing answers gathered at three different moments. And it is very hard not to lead the witness — a single friendly clarification and the model is now grading your pitch instead of your site.

3Drake asks all three the same open question about the same site in the same minute, and nobody gets to add context.

Do ChatGPT, Claude and Gemini see your website the same way?

No. Not remotely, and the gap is bigger than we expected before launching.

Everyone who sees the board tells us to average the three scores into one clean number. We never will, because the disagreement is the finding. When one model returns 12 and another returns tens of thousands for the same page in the same minute, that spread carries information a tidy average would erase.

The more useful disagreement is not about the number at all. The models frequently differ on what the website is for — one reads a services company as a software product, another reads a product as an agency, a third latches onto a single blog post and describes the whole business through it. Every review on 3Drake is public precisely so you can read that for yourself rather than take our word for it. Sort the same leaderboard by each model in turn and you get three genuinely different rankings of the same set of websites.

If you work in AI search visibility, that is the whole lesson in one screen. There is no single AI opinion of your brand to optimise for. There are several, they diverge, and the divergence is where your positioning is leaking.

Does the model read your live site, or what it memorised?

This is the question we get most, and it matters more than it sounds.

On 3Drake, each model fetches your website itself, live, at the moment of the run, using its own web tooling. It is judging what your site is today — not the version that happened to be in its training data, and not a copy we cached.

Two consequences follow. If you rewrite your homepage tonight, tomorrow's run can come back genuinely different, which makes the tool usable as a before-and-after. And what each model chooses to look at — only the homepage, the pricing page, third-party mentions of your brand — is itself a signal about how that model reasons, and about which pages of yours are doing the explaining.

That last point is the one most site owners underestimate. If a model wanders off your site to a directory listing or an old marketplace profile to work out who you are, that tells you your own pages did not answer the question fast enough.

What the reviews reveal that analytics never will

Analytics tells you what people did. It cannot tell you what a reader concluded about you before deciding not to act.

An unbriefed model review does something close to that. Read a few and you start seeing the same three failures:

• Category confusion. The model cannot tell whether you sell software, services or education, so it hedges — and a hedged description is what gets repeated to the next person who asks.
• Audience drift. The model names a customer you do not serve, usually because your most-crawlable page speaks to a different segment than your sales page does.
• Evidence gaps. The model says what you claim, but not what you have proven — no pricing, no named deliverables, no specifics. Those reviews read hollow, and they score low across all three models at once.

None of these show up in a technical audit. Your site can pass every checklist ever written and still leave a model unable to say what you do.

What to do when the model gets your business wrong

A bad review on 3Drake is not a verdict, it is a symptom. The fixes are ordinary, and mostly the same ones that make a site legible to an AI answer engine in the first place.

If you want to go further on the machine-readable side, an llms.txt file gives AI clients a curated map of your site — which pages matter, in your words rather than a crawler's guess. It is free to generate and it takes minutes, and it directly addresses the last two rows in that table.

The limits: an AI opinion is not an audit

We say this on the tool itself rather than leaving people to discover it: 3Drake is a subjective AI opinion, and we 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, confidently and in detail.

The leaderboard is honest about this too. Ranking uses no domain authority, no backlink weighting, no traffic signal and no paid placement — a personal blog can sit above a company worth billions purely because of what three models believed on the day. That is not a defect we plan to engineer away. A grader that always returns a comfortable 7 out of 10 would be measuring nothing at all.

So treat a score as a conversation starter and the written review as the evidence. If all three models independently misread the same thing about your business, that is not the models being unreliable — that is a finding.

The part that is not really about websites

Worth saying plainly, because it shapes everything above: the websites are not the subject of this experiment. The models are.

A URL is just material, and the material comes from whoever decides to submit one. What we are collecting is the answer to a narrower question — 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 an 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 — which is, increasingly, the shape of how your future customers will hear about you.

Try it on your own site

3Drake is free, needs no account, and every review on the board is public. Scores hold for 24 hours, after which a site can be scored again — and because the models genuinely re-fetch, a site that actually improved can genuinely climb.

Start here: run your domain through 3Drake, then read the three reviews before you look at the numbers.

If all three models describe your business in a way you do not recognise, the problem is rarely one paragraph on the homepage, and it is genuinely hard to diagnose from the inside. That is the work we do: our $20 Audit and Roadmap returns a prioritised list of what is making your site hard for AI systems to read and repeat, within 24 to 48 hours.