How HapHunt rates happy hours

HapHunt compares happy hour food prices to the same restaurant's regular menu prices, item by item. Nothing is compared across restaurants, and nothing is compared to a city average.

Only items that appear on both menus enter the calculation. For each matched pair we compute dollar savings and percent savings, then average the percent savings across all comparable food items to get the verdict.

That math currently runs on 3,926 matched item pairs across the fleet, drawn from 6,017 priced menu items. The median rated restaurant contributes 12 matched pairs, so a verdict is an average over about a dozen real price comparisons, not two or three.

No estimates. No vibes. Just math.

HapHunt only records menu items an average human eye could see and navigate to on the restaurant's own pages. If a guest standing at the door with a phone could not find the price, it does not become a price on our page.

Hidden structured data is never the sole source for an item. Plenty of restaurant sites carry machine-readable menus in their markup that no longer match anything on the visible page, and treating that as evidence is how a site ends up publishing dishes a kitchen stopped making.

PDFs, image menus, and menus behind a normal click all count as visible, because a guest can reach them. Third-party aggregator listings do not count as a source of record.

Every restaurant page links back to the source we read. 274 of the 276 active restaurants carry that link, so any price on this site can be checked against the restaurant directly.

Every item name and every price has to appear in the saved copy of the page it came from. An extraction whose name or price can't be found in that capture is routed out of the review queue entirely and sent back for a re-read, rather than being corrected by hand into something plausible.

Those captures are kept, not discarded. HapHunt holds 11,785 stored menu captures for the current Seattle fleet, which is what makes a price on this site auditable after the fact instead of merely asserted.

Every recheck re-reads prices item by item and records a per-item decision, so a change is attributed to a specific item rather than to a vague sense that the menu moved. Items a recheck could not confirm are marked unverified instead of being carried forward on faith.

Every restaurant page shows a “Last checked” date. That date is the day we last re-read the restaurant's own menu, not the day the page was built and not the day anything was published.

The target cycle is 30 days: HapHunt aims to re-read every published restaurant's menu at least once a month, with a nightly scheduler choosing which venues are due.

As of October 2026 that cycle is holding: 269 of the 276 active restaurants were last checked inside 60 days, and the median restaurant was last checked 12 days ago.

That has taken 6,848 completed recheck runs across 274 of the 276 restaurants. All 2 without a completed recheck were added in the last 30 days and are still inside their first cycle.

When a recheck finds the happy hour gone, the listing comes down. 22 restaurants have been archived off the live site that way. A directory that only ever grows is a directory that stopped checking.

If a detail isn't clearly listed, HapHunt labels it not listed rather than inferring it. An unlisted end time does not become “probably 6pm.”

We never publish a happy hour on a day the restaurant is closed. When a menu claims a window that lands on a day the restaurant's own hours say it is shut, the closure wins and a person has to reconcile the conflict before anything goes live.

A discount of 50% or more on an item whose portion size is never stated is treated as suspect, not as a win. Six wings for half the price of twelve is not a 50% discount, so 374 such items across 74 restaurants are currently held out of the averages instead of inflating them.

Items with no regular-menu counterpart are excluded from the savings math entirely. That is 2,091 items right now — a third of everything we have priced — and leaving them out is what keeps a happy-hour-only snack from being scored as a bargain against a price that never existed.

A restaurant with no comparable items gets no verdict rather than a bad one. We would rather say nothing than rate something we cannot compare.

Automated tools find and read the menus, and most routine re-reads close on their own. Human judgment is what the system is built around: Katherine Corson, who founded HapHunt, rules the calls the machine can't make, and each ruling becomes a rule the whole fleet then runs under. More about who runs this.

That autonomy is bounded, not assumed. Items disappearing from a menu and happy hour windows changing never apply on their own — those always escalate to a person. Anything else has to clear its evidence gates first: the new prices have to be found in the saved copy of the page, the re-read has to pass its own verification, and the scrape has to come back healthy. A nightly job then re-checks the published fleet for rule violations, so a listing that drifts out of compliance shows up the next morning rather than the next time someone looks.

The review that does happen is not a rubber stamp. Across 2,119 logged review sessions since April 4, 2026, 1,611 of them — 76% — changed something the machine had proposed. The grain there is review sessions, not restaurants; a single restaurant can be reviewed more than once.

Judgment calls become written rules rather than one-off fixes. The closed-day rule, the portion rule, and the visible-menu rule above all started as a single ruling on a single restaurant and now run against the whole fleet.

Happy hour menus change without notice. Source links sit on every restaurant page so you can check current prices and hours with the restaurant yourself — and if we have something wrong, we would rather hear it than not.