How we rank bars

We blend in 40 hypothetical 4.2-star reviews. Rating quality supplies 65% of the pillar. Review volume supplies 35%, with each additional review adding less and volume credit capped. “Clamp” limits a value to the 0–1 range.

smoothed = (rating × reviews + 4.2 × 40) / (reviews + 40); score = 100 × clamp(((smoothed − 3.5) / 1.5) × 0.65 + clamp(log10(reviews + 10) / log10(5,000)) × 0.35)

A bar may share a food-and-drink business with a restaurant or lounge. When we verify that relationship at the same location, we can use the shared listing’s rating. The venue page explains that it includes dining or other room guests and does not isolate bar-only visits. If two profiles use the same shared aggregate, both pages say so. We never use a whole-hotel rating or transfer a rating from another branch or an unrelated business.

Hong Kong and Macau stay within China on country rankings. Verified Google reviews can contribute to their guest rating. Macau accepts only newly reviewed exact-venue listings; older Google records stay historical. Mainland China’s Google records also stay historical and are never fetched for current research or scored. Google cannot establish current operating status in mainland China or Macau: those status checks still require an independent source. Previously reviewed locations remain intact.

For approved mainland-China and Macau estimates, an accepted Dianping observation takes priority over Baidu Maps. If neither is available, an exact-venue Tripadvisor listing takes priority over Trip.com or Ctrip. These are customer-rating fallbacks, not Google reviews. We then use the freshest observation from that provider, breaking an age tie in favour of the more reliable surface. We never average different platforms.

From 7 October 2026, separately reviewed exact-venue Tripadvisor, Trip.com and Ctrip fallbacks may also fill gaps outside China. Only the dated list of approved venues qualifies. Verified Google and local-provider ratings take priority. These travel-platform fallbacks always use low confidence and the 0.35 multiplier; they do not change the 23% Guest Ratings weight.

A proxy with review volume uses the formula above. Without a review count, its base is 50 × clamp((rating − 3.5) / 1.5). The base is multiplied by confidence: 0.85 high, 0.65 medium or 0.35 low. Hotel-wide, mall-wide, visit-count and unapproved parent records are excluded. Older observations affect confidence, not the point estimate.

Trip.com, Ctrip and Tripadvisor fallbacks always use the low multiplier, 0.35, even when read directly. Their customer samples can be small and differ from local review platforms. Each listing must match the exact bar and room; a hotel or sibling branch cannot supply its rating. A missing review count stays unknown. Guest Ratings still accounts for 23% of the overall formula.

We show whether a fallback came from a direct read or an indexed copy. The checked date records our source check, not when every review was posted or the aggregate last changed. An undisclosed cache age stays unknown. Version 3 adds these travel-platform fallbacks; existing Dianping and Baidu estimates keep version 2. Uncertainty ranges use a conservative allowance, not a calibrated statistical confidence interval: ±0.15 for direct travel-platform reads or ±0.20 for indexed copies, with another ±0.05 beyond each of six and twelve months, or both additions when age is unknown. A spread above 0.2 within the same provider adds ±0.05.

Number 1 starts at 100 points, number 50 at 51 and number 100 at 1. A result keeps 82% of its value for each year of age. Entries do not add together.

score = max(max(0, 101 − rank) × 0.82^age), for age ≤ 5

Top 500 Bars describes its method as an algorithm that gathers rankings, editorial coverage, platform reviews, social media and search data in more than 20 languages. We do not reproduce or alter that external algorithm; we score the published rank. Because some of its inputs can overlap with our other evidence, this pillar is limited to 10% of the total.

Number 1 starts at 100 points and number 500 at 0.2. A placement keeps 82% of its value for each year of age, matching our World's 50 Best decay rule. Editions do not add together. The age adjustment stays inside this 10% pillar and does not refresh the separate whole-score recency check. We add the 2026 edition in official published batches as each venue identity is checked. Published ranks 151–499 currently score; rank 500 is held for identity review. The publisher's intermediate page headed 101–150 exposed no venue rows, so those positions are not inferred. Ranks 176–181 keep their individual official-post sources and score once even though they also appear on the compiled 151–200 page.

The official 2025 pages currently show positions 1–450. Its 451–500 tab is empty, so we cross-checked those last 50 positions against a contemporaneous published transcription. Each placement links to the page used for that row.

score = max((max(0, 501 − rank) / 500) × 100 × 0.82^age), for age ≤ 5

Each win has a stored award weight of 1, 2 or 3. Multiple qualifying categories add separately. Items keep adding points, but each new item moves the pillar score less.

raw = Σ(award weight × 22 × 0.90^age), age ≤ 6; score = 100 × (1 − e^(−raw / 40))

Flagship programs add 15 points per stored award-weight unit. National programs add 12 and state or city programs add 8. Ranked results normally use weight 3 for numbers 1–10, 2 for 11–50 and 1 after 50. A source that proves list membership without preserving the exact rank receives the conservative weight of 1.

A bar can display every award it received, but only its strongest result from the same program and year earns points. For assessed guides, the latest accepted dated assessment scores once; a later reappraisal replaces the earlier status even when its rating is retained. It never stacks as another award. Annual results retain 85% per year. The Bartender's Edit uses its single strongest placement after volume-based decay: an older volume retains 85% for each edition behind that city’s latest volume. This is measured in editions, not calendar years. All raw points then enter the same diminishing-return calculation.

award raw = Σ(weight × (15 flagship or 12 national or 8 local) × 0.85^age), age ≤ 6; Bartender's Edit raw = max(max(0, 21 − rank) × 0.85^editionsBack × 5); score = 100 × (1 − e^(−raw / 50))

Duplicate URLs are normalized and counted once. Deduplication is within each venue. A link matching that venue's ordinary award evidence is also removed; known World's 50 Best and Spirited publisher sites are blocked separately.

If an editor-reviewed press record exists, only its approved links count. In code terms, its URLs form the exact allowlist, including an empty list that confirms nothing qualified. Without such a record, otherwise eligible cited coverage can pass.

Features, reviews, profiles and lists use multipliers of 1.00, 0.90, 0.85 and 0.45. Exact stored outlet labels receive these quality values: Punch, The New York Times and Michelin Guide 1.00; Imbibe and Financial Times 0.95; Difford's Guide 0.90; The Guardian 0.85; Bon Appétit, Food & Wine, Esquire and GQ 0.80; Wallpaper*, Robb Report, SCMP and South China Morning Post 0.75; The Infatuation 0.70; Time Out 0.55. Every other label receives 0.50.

raw = Σ(outlet quality × story-type multiplier × 18 × 0.80^age), age ≤ 6; score = 100 × (1 − e^(−raw / 50))

The research and synthesis workflow is proprietary. We publish the rule that affects the ranking: guest conversation supplies 70% and workplace treatment supplies 30%. Those shares do not move. If one read is missing, it contributes zero for its share. An estimate or invalid number is ignored, and the result is kept between 0 and 100.

Confidence, mention count, evidence count and update date do not change or decay the number. Venue pages show summaries, themes and source links when they are available. A source can also qualify as press; the two evidence sets are not deduplicated against each other.

Some older reads include guide or professional reviews. Our venue audits distinguish those source roles and sampling limits. “Score read updated” dates the stored number; “Guest sources checked” or “Staff sources checked” dates a source review. Checking a source again does not by itself recalculate the score.

Social Vibes = 0.70 × guest + 0.30 × staff; missing read = 0 for its share; neither read = Not Scored and contributes 0