The UX/UI score: how computer vision grades casino navigation
How GamblScout’s computer vision model scores casino website navigation, and what industry-wide UX and accessibility data reveal about the sector.
The Scoring System & Algorithmic Weights
We break down the scoring system and algorithmic weights behind GamblScout’s data-driven casino and sportsbook rankings.
GamblScout.com scores online casinos and sportsbooks with an algorithm, not a panel of paid reviewers. This category is the technical documentation for that algorithm: how it pulls data, how it reads player sentiment, how it weighs fraud and fair-play signals, and why we think that approach produces more reliable rankings than the sponsored "top 10" lists common across the industry. Below, we introduce the four sub-systems that make up the scoring engine and link to the detailed article covering each one.
Key takeaways
- Our scoring system rests on four pillars: a documented methodology, a data-scraping engine, NLP-based sentiment analysis, and fraud/fair-play verification — each covered in its own article in this category.
- US regulators have formally acknowledged the fake-review problem our approach is designed to avoid: the FTC's final rule against fake and paid reviews took effect on October 21, 2024, with civil penalties of up to $51,744 per violation.
- Courts in the hiQ Labs v. LinkedIn litigation established that scraping publicly available web data generally does not violate US federal computer-fraud law, though site-specific contract terms can still restrict it.
- Independent testing bodies such as eCOGRA, founded in 2003, already verify RNGs and payout data; our algorithm treats that certification status as one weighted input rather than a stand-alone guarantee.
- No single data point decides a score — weights are combined, cross-checked against regulator records, and revisited as new data comes in.
Table of contents
Why we score algorithmically instead of manually
Most casino and sportsbook "review" sites still work the way affiliate marketing has worked for two decades: a writer signs up, deposits a small amount, plays for an afternoon, and files a score that is heavily influenced by the commission the site earns if a reader clicks through. That model does not scale to a market of this size.
The global online gambling market size was valued at USD 99.7 Billion in 2025 and is to reach USD 179.7 Billion by 2034, at a CAGR of 6.80%
, spread across hundreds of licensed operators in dozens of jurisdictions. No small team of human reviewers can keep pace with that many license renewals, T&Cs updates, and payout-policy changes.
Regulators have also started treating manipulated reviews as a consumer-protection issue in their own right.
The Federal Trade Commission's final rule banning fake reviews and testimonials went into effect on October 21, 2024
, and the agency was explicit that
"AI tools make it easier for bad actors to pollute the review ecosystem by generating, quickly and cheaply, large numbers of realistic but fake reviews that can then be distributed widely across multiple platforms."
Violations can carry civil penalties of
$51,744 per violation
. That backdrop is part of why we built a scoring system that discloses its inputs rather than asserting authority through anonymous "our experts tested this" copy. Our full reasoning for abandoning the human-reviewer model is set out in Our Core Principles & The Problem with "Human" Reviews.
The four pillars of the scoring system
The algorithm is not one model but four connected sub-systems. Each has its own article in this category; the table below summarizes what each one contributes to a final score.
| Pillar | What it covers | Primary output feeding the score |
|---|---|---|
| Core principles | Why we replaced manual reviews and how conflicts of interest are structurally avoided | Governance rules that constrain the whole model |
| Data scraping engine | Automated collection of license data, T&Cs, RTP disclosures, payment terms | Structured, comparable operator records |
| NLP & sentiment analysis | Classification of player complaints, forum threads, and app-store reviews | A weighted sentiment and complaint-category score |
| Fair play & fraud detection | RNG/RTP certification status, security posture, and regulator complaint history | A trust and integrity score |
The remaining sections walk through each pillar and link to the full article covering it: Data Scraping & The Technical Engine, Natural Language Processing (NLP) & Sentiment Analysis, and Security, Fraud Detection & Fair Play.
Data inputs and scraping
Before anything can be scored, it has to be collected. Our crawlers pull licensing register entries, bonus terms, withdrawal-processing statements, payment-method lists, and RTP disclosures directly from operator sites and public regulator databases, then normalize them into comparable fields. This is the same category of activity litigated for years in the United States in hiQ Labs, Inc. v. LinkedIn Corp.
The Ninth Circuit court of appeals has held that data scraping public websites is not unlawful, affirming that plaintiffs may not rely on the Computer Fraud and Abuse Act to enjoin third parties from scraping data from their websites.
The case did not end there, however: after years of litigation the parties settled, and
the stipulation includes, among other things, a $500,000 judgment entered against hiQ and injunctive relief prohibiting hiQ's future ability to data scrape LinkedIn
based on breach-of-contract findings tied to specific conduct, not the CFAA itself. The practical takeaway we apply is that publicly posted operator data (licenses, published terms, disclosed RTP figures) is fair game for automated collection, while anything gated behind a login or a site's explicit anti-scraping terms is treated differently. The mechanics of our crawler, refresh cadence, and normalization pipeline are covered in Data Scraping & The Technical Engine.
NLP and sentiment weighting
Raw star ratings are close to useless for comparing operators, because they mix unrelated complaints (a slow customer-support reply, a rigged-game accusation, a delayed but ultimately paid withdrawal) into a single number. Our natural-language-processing layer instead classifies player-generated text — app-store reviews, forum posts, complaint-board submissions — into categories such as withdrawal delay, KYC friction, bonus-term disputes, and account restriction, and scores the sentiment within each category separately. This granularity matters given how much of the public review layer is contaminated by incentivized content; the same regulatory environment that produced the FTC's 2024 rule exists precisely because
fake reviews not only waste people's time and money, but also pollute the marketplace and divert business away from honest competitors
, in the FTC's own framing. A system that reads and categorizes raw complaint text at scale is less exposed to that kind of manipulation than one that simply averages star ratings. The full technical breakdown of our NLP pipeline, including how we detect and down-weight likely-incentivized reviews, is in Natural Language Processing (NLP) & Sentiment Analysis.
Fair play and fraud-detection signals
Sentiment tells us how players feel; it does not tell us whether a game is mathematically fair or whether an operator's security posture is adequate. For that, the algorithm leans on independent testing infrastructure that already exists in the industry.
Founded in 2003 and headquartered in the United Kingdom, eCOGRA is an independent game testing laboratory that specialises in the comprehensive evaluation, inspection, and certification of online gambling software and systems
, and it was
the first testing laboratory specialising in online gambling to have been awarded ISO/IEC 17021-1:2015 accreditation
. Certification status, RNG test results, and published RTP-verification reports feed directly into the fair-play score as a weighted input rather than a binary pass/fail.
We also track regulator-published complaint data as an external check on player-reported sentiment. UK Gambling Commission data obtained via Freedom of Information request shows
the total number of complaints concerning safer gambling, including complaints about self-exclusion, problem gambling, or gambling harm prevention failures, at 1,077 in 2024-25, 1,197 in 2023-24, and 1,135 in 2022-23
— a data series we use to sanity-check whether an operator's complaint trend is moving in the same direction as its scraped sentiment score. On the fraud side,
machine learning techniques can analyze historical data for betting histories, odds information, and players' gaming patterns to categorize anomalous betting patterns and irregularities that could indicate fraud
, according to legal analysis of AI use in regulated gambling — the same category of anomaly detection our fair-play pillar draws on when flagging outlier payout patterns. The complete methodology for this pillar is in Security, Fraud Detection & Fair Play.
How the weights are combined
None of the four pillars stands alone in the final score. Licensing and fair-play certification act as a gate: an operator with no verifiable license or no independent RNG certification is capped at a low score regardless of how positive its scraped sentiment looks, because that combination is itself a red flag. Above that gate, sentiment data (weighted toward the most recent 90 days of collected text, with older complaints decaying in influence) is blended with structural data points — withdrawal-time disclosures, KYC turnaround claims, payment-method breadth — pulled by the scraping engine. Fraud and security signals act as a further modifier: unresolved regulator complaint spikes or unusual payout-pattern flags reduce the composite score even when text-based sentiment is neutral. The exact numeric weightings are recalibrated periodically as we add data sources, which is why we describe the system as a set of pillars rather than a fixed formula — the relative importance of each pillar shifts as underlying data quality improves or degrades for a given market.
This layered approach is also why we treat the system as a complement to, not a replacement for, regulator oversight. Independent testing bodies and gambling commissions remain the authoritative source for licensing status and RNG compliance; our algorithm aggregates and cross-references that public information rather than substituting its own judgment for a regulator's.
Frequently asked questions
Does the algorithm replace human judgment entirely?
No. Human editors define the categories, review edge cases the model flags as ambiguous, and audit sample outputs for drift. What the algorithm replaces is the practice of a single reviewer's brief, subjective playtest deciding a rating — the scoring itself is computed from structured data rather than one person's impression.
How often is the underlying data refreshed?
Scraped structural data (terms, licenses, payment methods) is re-crawled on a rolling schedule tied to how frequently a given field tends to change; sentiment data is processed continuously as new reviews and complaints appear. Fair-play certification status is checked against the issuing lab whenever we detect a certification-related claim on an operator's site.
Can an operator pay to improve its score?
No input to the scoring model is for sale. This is a structural rule, not a marketing claim, and it is the direct reason this category exists — to document exactly which data drives a score so that the claim can be checked rather than taken on faith.
What happens when an operator's license status changes?
A license suspension, revocation, or jurisdiction change is treated as a high-priority signal that triggers an immediate re-scrape and re-score, rather than waiting for the next scheduled crawl cycle, because licensing status is one of the gating factors described in how the weights are combined.
Methodology note
This hub and its underlying scoring model draw on four signal families: structured data scraped from operator sites and public license registers; player sentiment extracted from reviews, forums, and app stores via NLP classification; independent RNG/RTP certification status from labs such as eCOGRA; and regulator-published complaint and enforcement data. Each family is refreshed on its own schedule and combined using the gating and weighting logic described above; none of it is influenced by commercial relationships with the operators being scored.
Gambling involves risk. Only play with money you can afford to lose and use the deposit limits and self-exclusion tools available in your jurisdiction.
