MFA detection, COPPA compliance, audience segmentation and carbon footprint — real scores from production endpoints, tested against URLs you already know.
Every sample set was generated with the same public API endpoints a customer calls. Each one ships three artifacts: a CSV with one row per URL and the key scores, a full JSON with the complete API response per URL for technical evaluation, and a one-page written summary of what the data shows. All of them are downloadable from this page, with no form in front of them.
50 URLs: 15 publicly documented MFA sites, 15 premium publishers, 10 mid-tier publishers, 10 deliberately borderline properties.
How MFA scoring works50 URLs across five buckets: child-directed, general child appeal, adult-directed, kids sites with data collection, and well-governed children's brands.
How COPPA scoring works50 URLs across 10 content verticals: personas, demographics, purchase intent, life stage and B2B firmographics.
How segmentation worksGrade any URL A–F for carbon intensity. Resource weight, third-party bloat, green hosting and optimization signals in one sustainability score.
How carbon scoring worksThe set is built to be adversarial rather than flattering: 15 domains publicly documented as Made-for-Advertising (Adalytics' 2024 "Made for Arbitrage" study and its network-graph clusters), 15 premium publishers that run heavy programmatic stacks and would break a naive ad-tech-count heuristic, 10 mid-tier regional and enthusiast publishers, and 10 borderline properties (celebrity, listicle and aggregation sites) where reasonable people disagree.
n = 15
78 mean score (min 78 / max 78)
n = 15
11 mean score (min 0 / max 25)
n = 10
15 mean score (min 0 / max 25)
n = 10
30 mean score (min 0 / max 79)
The headline separation: known MFA averaged 78 against a premium-publisher average of 11. Counting a premium publisher scoring above 45 as a false positive and a known MFA site scoring below 46 as a false negative, this sample produced zero of each. The borderline bucket is where the interesting behaviour lives: its 0–79 range is the model declining to give a single verdict to a mixed bag, which is the correct answer.
mfa_detection_sample.csv| Domain | Bucket | MFA score | Risk tier |
|---|---|---|---|
| heraldweekly.com | known_mfa |
78 | HIGH |
| kueez.com | known_mfa |
78 | HIGH |
| thedaddest.com | known_mfa |
78 | HIGH |
| trend-chaser.com | borderline |
79 | HIGH |
| suggest.com | borderline |
72 | HIGH |
| celebritynetworth.com | borderline |
29 | LOW RISK |
| boredpanda.com | borderline |
11 | CLEAN |
| dexerto.com | midtier_publisher |
25 | CLEAN |
| cleveland.com | midtier_publisher |
19 | CLEAN |
| cnn.com | premium_publisher |
18 | CLEAN |
| theguardian.com | premium_publisher |
23 | CLEAN |
| bbc.com | premium_publisher |
7 | CLEAN |
Read the borderline rows carefully.
trend-chaser.com and suggest.com were not on any public MFA list — they were
scored HIGH on their own structural and content evidence. That is the case the score exists for: the
arbitrage site nobody has published about yet. The full CSV carries every underlying signal
(ad script count, header-bidding depth, ad-to-content ratio, word count, clickbait, trustworthiness,
content originality, the public flag source, and the LLM's evidence sentence).
How MFA scoring works Try the live demo Download CSV Full JSON
Five buckets, chosen so that the sample can fail visibly: 15 unambiguously child-directed sites, 10 with general child appeal (gaming, UGC, wikis), 10 adult-directed business and technology sites, 10 kids' properties with real signup and data-collection flows, and 5 well-governed children's brands. Every page is rendered in headless Chrome with the consent dialog auto-accepted, so consent-gated trackers are actually observable.
n = 15 · mean risk 38 (2 LOW)
Child-directed detection fired on 15 of 15. Mean data-collection sub-score 2.3/20; mean tracking sub-score 10.2/20.
n = 10 · mean risk 27 (6 LOW)
Child-directed on 4 of 10. Data-collection 2.6/20; tracking 5.7/20.
n = 10 · mean risk 28 (5 LOW)
Child-directed on 0 of 10. Data-collection 3.3/20; tracking 10.3/20.
n = 10 · mean risk 31 (6 LOW)
Child-directed on 4 of 10. Data-collection 2.4/20; tracking 10.0/20.
n = 5 · mean risk 28 (3 LOW)
Child-directed on 5 of 5, but data-collection 0.6/20 and tracking 5.2/20 — the well-governed profile.
Child-directed sites carry adult-grade tracking loads (10.2/20 vs 10.3/20 on adult sites). Combined with is_child_directed=true, that is precisely the exposure a compliance buyer needs surfaced.
How to read it: the child-directed classifier is the anchor — it fires on every clearly child-directed site in the set and on zero adult sites. The risk score itself is deliberately not a verdict; it is a composite of five sub-scores (child-directed content, personal data collection, tracking and persistent identifiers, absent privacy controls, content appropriateness). Well-governed children's brands — policies that mention COPPA, light tracking — sit at the bottom of the range, exactly where they should.
coppa_compliance_sample.csv| URL | Bucket | Risk score | Level | Child-directed |
|---|---|---|---|---|
| www.coolmathgames.com | child_directed |
49 | MEDIUM | YES |
| www.abcya.com | child_directed |
47 | MEDIUM | YES |
| www.mathplayground.com | child_directed |
44 | MEDIUM | YES |
| www.poptropica.com | child_directed |
40 | MEDIUM | YES |
| poki.com | general_child_appeal |
40 | MEDIUM | YES |
| www.roblox.com | general_child_appeal |
33 | MEDIUM | YES |
| outschool.com/signup | child_data_collection |
42 | MEDIUM | YES |
| www.getepic.com/signup | child_data_collection |
36 | MEDIUM | YES |
| tocaboca.com | good_compliance |
20 | LOW | YES |
| techcrunch.com | adult_directed |
44 | MEDIUM | no |
| www.marketwatch.com | adult_directed |
20 | LOW | no |
How COPPA scoring works Try the live demo Download CSV Full JSON
50 URLs across 10 content verticals — luxury, B2B enterprise, parenting, finance, health, automotive, tech and gaming, food, fashion and beauty, education. All 50 returned full audience profiles. Every URL combines static persona mapping (which costs no LLM call) with LLM-inferred demographics, purchase intent, life stage, B2B firmographics and content context.
5/5 enterprise-content URLs classified is_b2b=true.
4/5 URLs assigned a high / affluent income level.
50/50 URLs returned a full profile — no gaps to explain away.
audience_segmentation_sample.csv| URL | Vertical | Top personas | B2B | Top purchase intent |
|---|---|---|---|---|
| robbreport.com/motors/cars/ | luxury_affluent |
Exotic Car Enthusiast; Luxury Car Enthusiast; Auto Enthusiast | B2C | luxury vehicles (0.90) |
| www.hodinkee.com | luxury_affluent |
Minimalism Enthusiast; Fashion Enthusiast; Organic Food Advocate | B2C | luxury watches (0.90) |
| www.snowflake.com/en/blog/ | b2b_enterprise |
Infrastructure Engineer; Data Scientist; Cloud Computing Specialist | B2B | technology (0.80) |
| hbr.org | b2b_enterprise |
Financial Analyst; Corporate Executive; Financial Advisor | B2B | business_management (0.85) |
| stackoverflow.blog | tech_gaming |
Technology Enthusiast; Cloud Computing Specialist; Data Analyst | B2B | software tools (0.80) |
| www.whattoexpect.com/first-year/ | parenting_family |
Pediatrician; Childcare Professional; Parenting Blogger | B2C | baby products (0.90) |
| www.fool.com | finance_investing |
Cryptocurrency Investor; Beginner Investor; Financial Analyst | B2C | investment services (0.90) |
| www.healthline.com | health_wellness |
Health Enthusiast; Exercise Enthusiast; Detox Devotee | B2C | healthcare products (0.80) |
| www.ign.com | tech_gaming |
Gamer; Action Enthusiast; Hardcore Gamer | B2C | video_games (0.90) |
| www.harpersbazaar.com | fashion_beauty |
Fashion Forward Female; Fashion Enthusiast; Style Enthusiast | B2C | fashion (0.90) |
The CSV carries considerably more per URL than fits in a table: age band, gender skew, income level, education, life stage, B2B roles, content type, reading level, price signals and IAB categories. The JSON carries the complete response, confidence scores included.
How segmentation works Try the live demo Download CSV Full JSON
Grade any URL A–F for carbon intensity. The score (0–100, lower is greener) combines resource weight analysis, third-party script bloat, green hosting detection, image optimization and render efficiency into a single sustainability metric. Six grade tiers: A (0–15), B (16–30), C (31–50), D (51–70), E (71–85), F (86–100). Use it for sustainability reporting, green advertising programs and brand placement decisions that factor in carbon cost.
Total page weight, JavaScript payload, CSS payload, image weight and font loading.
Number and weight of third-party scripts, ad-tech load, tracking overhead and render-blocking resources.
CDN usage, hosting provider green energy status, HTTP/2 adoption and caching efficiency.
A sample report that only shows what worked is a sales asset, not evidence. These are the caveats we would raise ourselves in front of a media auditor, and they are stated in the README that ships with the data.
Several documented MFA operators serve ad-free, harmless-looking pages to datacenter IPs. The monetization_observed field reports what our crawler actually saw, kept separate from the domain flag — a flagged domain showing no ads to us is a cloaking fingerprint, not innocence.
A single-fetch API cannot walk a multi-step JavaScript registration wizard. Those pages report signup_cta_detected rather than an enumerated list of PII fields. We report the CTA; we do not claim to have completed the form.
Every URL was live and returning HTTP 200 with real content at generation time (2026-07-13). MFA domains turn over month to month — re-verify liveness immediately before any live demo, and re-score inclusion lists on a schedule.
COPPA pages are rendered in headless Chrome with CMP auto-accept. Crawling the EU without accepting consent hides most of the tracking stack, which would make every site look cleaner than it is.
Bucket assignments (known MFA, premium, borderline) come from public research and editorial judgment, not from the model. That is what makes false positives and false negatives measurable in the first place — and it is also why you should check the domains you know personally.
URLs owned by top-100 advertisers were excluded from the sample sets to avoid conflicts of interest in an audit context. Scores are produced by the same public endpoints a customer calls — no offline tuning, no per-domain hand-editing.
Each demo scores a live URL in real time using the same production endpoints behind the sample data above. No signup, no API key — just paste a URL and see the result.