We audited ourselves and failed.
We sell the detection of generic AI-generated design patterns. Our own marketing site broke several of the standards it sells. This is the score before, every finding, what we changed, and the score after — measured with our own scanner and the rubric published at /methodology.
The numbers
Lower is better. Both snapshots were scored by the same deterministic path, with the model pass disabled, so the comparison is arithmetic rather than opinion. “Before” is the deployed homepage; “after” is a production build of the remediated one.
| Dimension (API field) | Before | After | Change |
|---|---|---|---|
| visualOriginality | 95 | 95 | no change |
| copyOriginality | 73 | 73 | no change |
| typographyDiscipline | 96 | 96 | no change |
| layoutAsymmetry | 96 | 96 | no change |
| brandVoice | 79 | 79 | no change |
| Detector | Before | After |
|---|---|---|
| Radial gradient glow | clear | clear |
| Glassmorphism panel | clear | clear |
| Fully rounded pill button | fired | fired |
| Sparkle icon | clear | clear |
| Three-column card grid | fired | fired |
What the score does not capture
This is the uncomfortable part, and it is the reason the post is worth reading. The score did not move. Not one of the five dimensions changed. We rewrote or deleted every problem listed below, and our own scanner reported an identical result before and after.
The scanner cannot see an unsourced statistic. It has no detector for an invented taxonomy, for four fabricated people with quoted opinions, for a technical term that describes something the product does not do, or for a published weight table that does not match the engine. Those were the most serious problems on the site and every one of them was found by a person reading the page against the standard we sell, not by the tool.
We also found two false positives in our own favour and against it. Our copy scores a cliché penalty for the phrase “empower your workflow” — which appears on our site only because we quote it in order to criticise it. And while writing the methodology page we discovered that documenting the detectors makes them fire: an early draft of the FAQ listed the exact class names each detector matches, and the scanner dutifully flagged our own documentation as a purple-gradient site. The detector matches strings anywhere in the document and never checks whether the match is visible. We reworded the FAQ and moved the literal strings to /methodology, which means that page would score badly if you scanned it. That is a real limitation of the product, it is now item three on our published list, and we would rather you heard it from us.
Two detectors still fire on our homepage and we are leaving both. The three-column grid detector is correct: we do use three-column rows, and a grid is not a defect. The pill-button detector is a false positive — every match is a six-pixel status dot, and the detector cannot tell a dot from a call to action. The cliché penalty is the “empower your workflow” quotation described above. We would rather publish a score of 18 with the reasons than chase a lower one by deleting a status dot.
A low slop score is not proof of craft. It means the detectors we have did not fire on the markup we could see. That is a narrower claim than it looks, and anyone selling you a design score should tell you so.
Every finding
01Three quantitative claims with no source
“3.4× Higher ACV”, “−58% Bounce Drop”, and “Stop burning 64% of traffic in the hero viewport”, plus a “refactor UI in 15 minutes” and an unsourced “40% to 65% higher bounce rates” in the FAQ. Our own copyOriginality dimension penalises unfounded claims and rewards verifiable ones. These were worse than vague marketing language, because they were specific enough to be checked and could not be.
Deleted, not softened. The three commercial blocks now argue from mechanism. In their place we publish figures computed from the sites we have actually scanned, each carrying its sample size and the date it was computed.
02A decorative identifier taxonomy
PHASE // 01, DOM-AST-01, HEURISTIC-02, REMEDY-03, DIM-01·DOM-AST-01, FAQ // 01, and numbered jury personas. Structural markers used as ornament, encoding nothing. This is a pattern we flag on other people’s sites.
Removed, except where the marker carries information: the 01/02/03 on the pipeline stayed, because crawl-then-score-then-remediate is a real sequence. The five dimension labels stayed but were corrected to the field names the API actually returns, so a card on the page maps to a field in your response.
03Emoji throughout a site that bans them
Phone, tag, lightning, ladybird, clipboard and others across the homepage, pricing, extension page, footer and toasts.
Removed from the marketing site and every user-facing string. Kept in three places: the mockups that demonstrate the pattern being criticised, the detector that matches on the sparkle glyph, and internal operator tooling.
04“AST” describing something we do not do
DOM-AST Sweep, AST scanner, computed AST token decomposition, JSON AST diffs. We do not parse an abstract syntax tree. We extract computed properties and text from a document — which the FAQ described correctly while the homepage contradicted it.
Replaced across 34 files with the correct term in each context. The two references to ESLint AST rules stayed, because ESLint genuinely does parse one.
05A determinism claim that could not be true
“Deterministic AST Rules · Zero Guesswork” sat on the same page as “Art Director Deep Reasoning”, which is a language model. A gauge footnote read “100% DETERMINISTIC”.
The claim now splits along the real seam: extraction is deterministic and reproducible, scoring combines fixed heuristics with a model pass for the critique and the buyer jury. Naming which half is which is a selling point, not a weakness.
06Four invented people with fabricated quotes
The buyer jury presented Alex Chen, Elena Vance, Marcus Reynolds and Jordan Miller with trust scores and quoted reactions. Fabricated named individuals is a recognisable model output shape, on a page selling the detection of recognisable model output shapes. It also read as fake testimonial at a glance.
The roles remain, the people are gone: staff infrastructure architect, finance buyer, design critic, five-second visitor. Both surfaces now state in the interface that these are model-generated evaluations, not real people. The model prompt was changed to forbid inventing a name.
07A published weight table for a weighting that does not exist
The methodology drawer published Visual 25%, Copy 30%, Typography 20%, Layout 15%, Brand 10%. A second component published a different set. Neither exists in the engine: the five dimensions are computed independently and never combine into the headline score.
Both deleted. /methodology now publishes the real arithmetic, the exact strings each detector matches, and an eight-item list of what the scanner cannot see.
08Benchmark scores for named companies that we had not measured
The hero showed “Stripe 96%” while our own scan of stripe.com scored it differently, and two of the three named companies had never been scanned at all — their numbers were hand-authored. The three chips also mixed two different scales in one row.
Every benchmark now comes from a real scan in our corpus, on one consistent scale, dated, and linked to the methodology. A domain we have not scanned cannot appear.
09Smaller credibility leaks
“All systems operational” hardcoded to no health check. “Unlimited URL audits (Fair-Use*)” where the asterisk resolved to a 150/month cap. A follow-for-credit growth hack in the pricing area. “SlopCheckr Forensics Lab” implying a lab.
The status string is gone. The tier says 150 audits per month. The growth hack moved to the footer. The author is now just SlopCheckr.
Reproduce it
The numbers above are generated, not typed. The script fetches both snapshots, scores them through the same engine with the model pass disabled, and writes the JSON this page renders.
npx tsx scripts/self-audit.ts before.html after.html src/data/selfAudit.json
Scored by SlopCheckr Multi-Section DOM Inspector v2.5 (Algorithmic Pass) on 31 August 2026.
Run the same scan on your own site
Same scanner, same rubric, no account needed for the first one.
Audit a URL