When people ask how a tiny team keeps a 7,000-game catalog clean, the assumption is always the same: "you can't — it must be automated." It is, partly. But the automation only handles the obvious cuts. Every borderline title is reviewed by a person, and every game promoted to a best of list or guide has been played at least once by someone on the team. This post is the long version of how that pipeline runs, what gets cut and why, what the math behind our rankings actually looks like, and the mistakes we have made along the way.
We have wanted to write this for a while because curation in this space is mostly invisible. Players see the games that made it. They don't see the ones we threw out, the categories we collapsed, or the ranking formula we use to decide what shows up first. Pulling the curtain back is partly transparency and partly an invitation: if you spot a title on Glaze Games that fails one of the bars below, tell us — that means the pipeline missed it and we want to know.
#The pipeline in one paragraph
A new title enters our system from one of three licensed HTML5 distributors, gets normalized into a single catalog schema, passes through a chain of automated filters, lands either on the live site (clear yes), in a quarantine queue (clear no), or in an editorial review pile (everything in between). A reviewer hand-plays anything in the review pile and marks it live, rejected, or "kids-safe only." Once live, the title is ranked against everything else in its category by a single visible formula. Nightly, the entire process repeats and any title that has degraded — broken embed, new ad behavior, plummeting rating — gets pulled. That's the whole loop.
#Step 1 — Nightly intake & API normalization
New titles come in nightly from licensed HTML5 publishing partners. Each distributor ships metadata in a completely different shape — categories that do not align, controls described in free-text, and ratings measured on entirely different scales. The intake step rewrites all of it into a single schema: a slug, a canonical title, a one-line description, a primary category drawn from a fixed list of about thirty buckets, a tag set drawn from a controlled vocabulary, an orientation, a controls fingerprint (keyboard, mouse, touch, gamepad), a hosted iframe URL, an aspect-ratio hint, and an integer play count.
Why three distributors and not one? No single HTML5 distributor has everything. Some partners are strongest on casual and Puzzle games, others are broader across arcade, simulator, and .io multiplayer arenas, and others lean toward the kids-and-family end. Pulling from multiple sources gives us a fatter catalog and a way to spot quality differences — a game that all three carry under the same name is almost always a known good. A game that shows up only in one tier-three distributor with zero active plays is almost always something we shouldn't bother indexing.
#Step 2 — Automated filters & hard-no rules
Before any title reaches our editorial review team, it must pass through a strict, automated screening pipeline. This pipeline uses headless browser simulations (Puppeteer) and network interceptors to analyze game payloads. We run these simulations across five common viewport resolutions (desktop 1440px, laptop 1280px, tablet 1024px, large phone 414px, and small phone 360px) to verify responsive scaling.
#The Hard-No Rules (Instant Auto-Quarantine)
- Age Ratings: Any title with an age rating above 12+ is immediately excluded. We aim to keep Glaze Games fully safe to leave open in any household.
- Gambling & Monetization Dark Patterns: Any game flagged with gambling, casino, real-money betting, or "lootbox-as-a-core-mechanic" tags is permanently banned.
- External Redirections: Any title that triggers a prompt to download an
.apkfile, redirects to external app stores, or spawns off-site popups within the first 60 seconds of load is flagged as malicious. - Audio Behavior: Any title running auto-playing audio without a visible, programmatically accessible mute control fails our compliance check.
- Network Security Blocklist: We intercept network requests to block games associated with aggressive ad networks that serve popunder ads, force malicious browser history rewrites, or execute cross-site tracking.
- Permission Harvesters: Any casual browser game requesting invasive permissions — such as geolocation, camera, microphone, clipboard-write, or push notifications — is blocked.
#The Soft-No Rules (Route to Editorial Review Pile)
- Poor Quality Scores: The quality rating falls below the 45th percentile for its respective category.
- Missing Documentation: The game provides zero instructions text and lacks an in-game tutorial sequence.
- Suspicious Tag Overlap: Overlapping tags that conflict logically (e.g., tagging a game as both a 2 Player chess game and an FPS shooter).
- Anomalous Play Count Growth: The play count experiences a sudden 50x spike without any corresponding increase in natural ratings, indicating potential botting.
#Step 3 — Human Editorial Review
Borderline titles, games in our Safe for Kids pipeline, and every single game promoted to a best of list are played by our editorial team. Reviewers are not asked to score subjective "fun" — instead, they grade each game against an objective quality rubric.
Reviewers run the game on an entry-level school Chromebook to guarantee performance on low-end hardware, checking if the title achieves a stable 30+ FPS over 3 minutes of play. They verify that the controls are intuitive and responsive across both desktop (keyboard/mouse) and mobile (touch targets). They also confirm that no intrusive ad behavior, such as non-skippable mid-game video overlays, occurs during the initial session.
If there is a disagreement, we route the game to a second editor. If they also vote "no," the rejection is final. If they disagree, the game enters a 24-hour cooling-off period before a third editor breaks the tie. We log every single rejection reason programmatically to keep our team's quality standards fully calibrated and consistent.
#The Mathematics of /best Lists
Every ranking page on our site uses a single, fully transparent formula. We do not accept paid placements, and we do not alter rankings to favor specific publishers. The algorithmic ranking score is calculated as follows:
PopularityScore = PlayCount + (AverageRating * 1000)
Why do we combine these metrics? If we ranked games solely by rating, a new game with only 5 five-star reviews would sit at a perfect 5.0, dominating the list over a beloved classic. Conversely, if we ranked purely by play count, older viral hits would permanently hog the top spots, blocking newer, high-quality titles from getting discovered. By multiplying the average rating (scale 0-5) by 1,000, we make one full rating point worth exactly 1,000 plays. This allows an outstanding new game with a 4.8 rating and 5,000 plays to easily outrank a stale 1,000,000-play game whose rating has slipped to 3.2.
To protect our rankings from manipulation, we require a strict minimum threshold of 250 verified user ratings before any game becomes eligible for a Best-of list. This completely neutralizes rating-injection bots and guarantees that our recommendations reflect genuine, stable player satisfaction.
#The .io/Clone Epidemic & Asset Fingerprinting
The HTML5 gaming ecosystem is heavily flooded with reskins — developers buy generic templates for popular genres (like Clicker or .io games), swap a few sprites, and submit them as new games. To combat this, we built an asset-fingerprinting system. Our system analyzes canvas rendering signatures, audio asset hashes, and source-code structures. If a submission shares more than 85% of its code and assets with an existing game, our pipeline flags it as a clone, ensuring we only catalog games that offer a genuinely distinct experience.
#Nightly Cleanup & Catalog Longevity
Curation is not a one-time event; it is a continuous loop. Every single night, our automated system runs health checks across our active catalog. It flags any embed URL returning a 4xx or 5xx error, any game whose rating drops below 2.5, or any title that has updated its SDK to pull in restricted ad networks. On average, we remove 200 to 400 degraded games every night, ensuring that Glaze Games remains the cleanest, most reliable portal on the web.