How we rank

Our coffee-shop list uses a calculated composite score. Software combines the available evidence using fixed rules. An AI model does not choose the order.

The current coffee pilot covers a limited set of venues with usable recent evidence. It is not a survey of every café in Dublin. Other categories may still use earlier data and methodology versions.

The coffee scoring formula

Each component is scored from zero to one. The final score is their weighted sum. We do not stretch the results to give the winner a perfect score. Category pages show the weights used for their published list.

Ratings and missing providers

The coffee formula assigns reputation weights of 55% to Google, 30% to Tripadvisor and 15% to Yelp. It uses only available valid observations and redistributes those weights within reputation. When Google and Tripadvisor are available, their shares are approximately 64.7% and 35.3%. Yelp is currently unavailable in this pilot.

Averages from small samples are adjusted towards that provider’s category average using a 50-review prior. Review counts influence this adjustment rather than giving popular venues an unlimited bonus. Different providers have different audiences, so their ratings are imperfect comparisons.

The coffee list requires at least two verified sources per venue, including a recent customer-rating input. One review platform can contain hundreds of customer reviews. A second source may be another review platform or an independent coffee guide. Missing evidence is not treated as a bad customer rating.

Guides and venue matching

Recommendations contribute according to publisher weight, recommendation type and age. Each publisher contributes at most its strongest recommendation for a venue. Repeated copies of the same article do not add votes. Paid guide profiles are not counted as independent endorsements without separate supporting evidence.

We match individual locations using provider identifiers, street addresses, closely matching coordinates or explicit map references. A shared brand name is not enough to combine two branches. Ambiguous matches are excluded. Unknown article dates receive reduced weight and do not become recent coverage merely because we fetched the page again.

Trend and history

Trend uses dated coverage and changes in customer-review counts. We save private observation snapshots so later runs can compare the same venue and platform. Review growth needs observations at least seven days apart. Until there is usable history or recent dated coverage, the breakdown says “Not enough history”. An unavailable trend contributes no bonus, and the other component weights stay fixed. We do not invent earlier counts or treat fetching an old article as new coverage.

Weekly updates

The coffee job runs each Monday at 06:23 UTC. It gathers evidence, calculates the list and checks the result before requesting a website rebuild. New entries, large movements, stale ratings or integrity failures can hold an update. The first list establishes the baseline for later comparisons.

If checks or deployment fail, the previous successful website remains available. The date on a ranking page is its ranking release date, not the date the website was built.

Limits

A category guide may also show clearly labelled editorial inclusions below its ranked list. These have no numerical rank or composite score. A business website can confirm its services, but does not count as an independent review source. An editorial inclusion moves into the calculated ranking only when it meets the same evidence requirements as the other candidates.

Coverage varies between venues and changes as sources become available. Publisher weights and the scoring parameters are modelling choices, not scientific proof that one café is better than another. Our scores and ordering are our own composite outputs. We do not display the underlying providers’ star ratings on business pages.