Teardowns / B2B SaaS

Three SaaS pages, roasted

All three founders asked for a roast in public, on X, in May 2026. All three built something genuinely good, then opened the page on a line written for themselves instead of the buyer. Same product, same week, same failure mode.

Roast Nº 01 · LLM cost observabilityrequested by @toflou, 29 May
Tokenwise tokenwisehq.com

The product watches every LLM call from one line of code and tells you where you are overpaying. The proof is real: $4,128 saved across 48 teams last month, 14.8M tokens analysed, 96.4% quality match. The hero leads with none of it. It opens on the mechanism, "One line.", and buries the money below the fold.

Before
One line. Every LLM call, watched and optimized.
Observe every call, catch every regression, and cut up to 30% off your LLM bill.
CTA: Try for free
After
Cut your LLM bill by up to 30%, from one line of code.
$4,128 saved across 48 teams last month, with a 96.4% quality match so you never trade output for savings.
CTA: See my LLM spend in 5 minutes

What is breaking

  1. Mechanism over outcome: "One line." describes how it works, not the smaller bill I actually buy.
  2. Proof below the fold: the $4,128 counter is the most persuasive thing on the page and a first-time visitor never reaches it.
  3. Three jobs, two buyers: observe, catch regressions and cut cost sells the engineer and the founder in one breath.
  4. CTA names no reward: "Try for free" asks for a trial before showing the first session's payoff.
  5. The 96.4% is buried: the one number that proves you cut spend without degrading output sits in a stat strip.

The fixes, ranked by ROI

  1. Lead with the bill: the outcome a founder feels in the bank goes on line one.
  2. Pull one number up: "$4,128 saved across 48 teams" does more in the hero than a wall of testimonials below.
  3. One buyer per line: money buyer in the hero, engineer gets their own line right under it.
  4. CTA previews the payoff: "See my LLM spend in 5 minutes" keeps the one-line promise and shows the reward.
  5. Let 96.4% breathe: frame the fear it removes, "what do I lose by cutting cost", right near the hero.
The full roast is publicTokenwise was the first field note. You can read the complete teardown, with the hero rewritten and ready to paste, on its own page. Read the Tokenwise roast in full →
Roast Nº 02 · multi-agent orchestrationrequested by @marcokueks, 27 May
Struere struere.dev

Real depth here: multi-agent orchestration across 40+ models is rare. But the hero opens on "AI-native companies run on closed loops," a Patrick-McKenzie-grade concept dropped before you have shown what the product does. You have to already get the idea for the line to mean anything, so people bounce before the value lands.

Before
AI-native companies run on closed loops.
Use cases: Recruitment · Property ops · Clinic. (shown as labels, no screenshots)
CTA: Get started
After
Your support agent handles a complaint, updates the CRM and books the call-back. On its own.
That is a closed loop: agents that act, check the result and improve. Across 40+ models, in one place.
CTA: Watch one loop run

What is breaking

  1. Concept before demonstration: "closed loops" is the label, not the thing. It needs the scene first.
  2. No felt scenario: nothing shows the six seconds of work the product actually does.
  3. Use cases as vapor: Recruitment, Property ops, Clinic shown as labels with no screenshot reads like vaporware.
  4. Depth hidden: 40+ models and orchestration, the real moat, is not in the first screen.
  5. Generic CTA: "Get started" asks for commitment before the visitor understands the payoff.

The fixes, ranked by ROI

  1. Open with a six-second scenario: agent handles a complaint, updates the CRM, books the call-back.
  2. Then name it: call that a closed loop once the reader has felt it.
  3. One real screenshot per use case: a frame under each label shifts it from claim to proof.
  4. Surface the moat early: 40+ models and orchestration belong near the hero, not three scrolls down.
  5. CTA shows the product: "Watch one loop run" previews the magic in one click.
ResultA visitor who did not already know the term now watches the product do a job in six seconds, then learns the word for it. Concept earns its place instead of gatekeeping the page.
Roast Nº 03 · AI app builderrequested by @thehenryinsf, 17 May
Blink blink.new

The hero claims "Blink is the world's most complete AI-powered full-stack app builder." "World's most complete" trips the bullshit detector, and paragraph two opens by listing seven competitors Blink is not. Listing eight rivals in P2 reads defensive, and the 1M+ stat lands before the visitor knows what Blink does for them.

Before
Blink is the world's most complete AI-powered full-stack app builder.
Blink is different from Cursor, Copilot, Lovable, Bolt, Bubble, Webflow, Replit. 1M+ builders, 3M+ apps built.
CTA: Start building
After
Describe an app. Ship a working version in 10 minutes.
Then keep editing it in plain English, frontend, backend and database together. 3M+ apps already built this way.
CTA: Build my first app

What is breaking

  1. Unprovable superlative: "world's most complete" means nothing to a visitor and triggers doubt.
  2. Defensive P2: naming seven competitors plants rivals the reader had not even considered.
  3. Stat before meaning: "1M+ builders" lands before I know what the thing does for me.
  4. "Complete" is empty: the word carries no concrete promise a builder can picture.
  5. CTA is a shrug: "Start building" names no first outcome.

The fixes, ranked by ROI

  1. Lead with one shipped app: "ship a working version in 10 minutes" is concrete and testable.
  2. Drop the competitor list: stop arguing against tools the visitor was not thinking about.
  3. Move proof after the promise: 3M+ apps means something once the visitor knows the job.
  4. Name the mechanism plainly: edit in plain English, frontend, backend and database together.
  5. CTA names the reward: "Build my first app" previews the ten-minute payoff.
ResultThe reader pictures a finished app in ten minutes before any stat or rival is mentioned, which is the only thing that makes the 3M+ number persuasive instead of suspicious.

Your SaaS page probably leaks in the same place.

Send me the link. I write back this exact breakdown, with your hero rewritten and ready to paste, within two hours for €25.

Roast my SaaS page, €25
Other teardowns: coaches and consultants · shops and agencies