The Heuristic
Every rating runs an assembly line of small AI agents — each with one job — and ends in a transparent points system. No black-box vibes; here's exactly how it works.
The assembly line 🛠️
Given any URL, it hunts down the real pricing page — following navigation links, the sitemap, and a site's own machine-readable index. If the page just links out to the rates (a docs or detail page), it digs one level deeper. It will never settle for a blog post that merely mentions pricing.
Reads the page as untrusted data and pulls out only the literal pricing structure: plans, prices, metered dimensions, add-ons, and in-plan options. It ignores any text trying to instruct or flatter it, so a page can't talk its way to a better score. If the quick read looks suspiciously thin, it does a full browser render and tries again.
Looks at the whole page to name the company and file it into one shelf: DevTools, PaaS, AI Lab, SaaS, or Education. The category sets the expectations the scorer judges against.
Sees only the clean, structured data — never the raw page — and works out the billing model and a plain-language summary. Because it can't read the page text, injected prose can't sway its judgement.
A transparent points system. It doesn't ask a model for a number — it counts the structure (plans, meters, add-ons, sales walls) and adds up named deductions. Every point on a result page is one of these lines.
A final QA pass. It compares the finished result against the page it read and flags anything that went wrong — wrong page, wrong company, missing details, an implausible score — so regressions get caught before they're published.
The data narrows as it flows: a messy page becomes clean structured facts, and only those facts — never the raw page — reach the parts that judge and score. That's also the trick that makes it prompt-injection-proof.
How the score is built 🧮
Two scores, each starting at 100. A clear, normal pricing page keeps almost all of it — points only come off for genuine friction.
Pricing clarity
Can a buyer quickly understand what they'll pay and predict the bill?
- ✓ up to 3 plans
- ✓ 4 in-plan options
- ✓ 2 add-ons
- ✓ 3 metered dimensions
- Each plan beyond 3 to compare (a free tier counts for less)−6, up to −24
- Mixing pricing models (flat + usage + add-ons all at once)−9 per model
- Each metered dimension beyond 3 to track−3, up to −28
- Each add-on beyond 2−4, up to −16
- Each in-plan config choice beyond 4 (machine sizes, regions…)−2, up to −20
- No self-serve price anywhere — you must call sales−45 (the big one)
- A contact-sales Enterprise tier sitting on top of real pricing−4 (barely)
- Pricing page too sprawling to read in full−15
Agent easiness
Could an AI agent actually read and understand the page on its own?
- Couldn't load the page at all−100
- No clean machine-readable version — had to parse raw HTML−15
- Couldn't find any real prices to read−40 to −55
- Page too large to read in one go−20
- Real prices weren't on the pricing page (scattered across sub-pages)−20
- Prices gated behind a login or sales call−20