The venue was fully staffed on Fridays. That was the confusing part, and it is why the problem had survived four months of being complained about.
Twelve of the previous nineteen negative reviews were written about a Friday. Of those twelve, nine described a visit that began between 8pm and 9.30pm. The language was consistent and slightly odd: not "the food was bad" or "the server was rude", but "felt forgotten", "had to flag someone down twice", "everyone seemed busy but nobody was looking at us".
Nine reviews saying the same unusual thing is not a run of bad luck. It is a finding with a time stamp on it.
A composite case, with representative numbers. The diagnostic sequence is the transferable part.
Staffing level and staffing shape are different problems
The Friday rota had eight front-of-house staff, which was the highest of the week and correct for the covers. But the shifts were built as two blocks: a 4pm–11pm long shift and a 6pm–close shift, with breaks taken when the floor allowed.
In practice "when the floor allowed" meant around 8pm, because that was the lull between the early sitting leaving and the late sitting arriving. Three people went on break at roughly the same time, every Friday, because that is when it felt quiet.
It felt quiet because the early tables were paying and the late tables were not yet seated. Twenty minutes later, four tables of six walked in at once, and the floor was running at five staff instead of eight with three sections uncovered — and would stay that way until the breaks finished at around 8.40pm.
The reviews were being written by the people seated in that window. "Everyone seemed busy but nobody was looking at us" is an exact description of a floor running short by three.
The check that found it in twenty minutes
Nobody needed new software for this. The sequence was:
- Filter reviews to negatives only, then group by day of week. Fridays were 63 per cent of them against 19 per cent of covers.
- Read the reviews for an arrival time, which guests mention surprisingly often — "got there about half eight". Nine of twelve gave one.
- Put those times next to the rota and the break log. The overlap was immediate and total.
The third step is the one that gets skipped. Reviews on their own tell you when service degraded. They cannot tell you why unless you lay them against something that describes what the building was doing at that hour: the rota, the break log, the booking sheet, the section map. The finding lives in the overlap, not in the reviews.
What changed, and what it cost
Breaks were staggered in three waves from 7.15pm, with a rule that no more than one person is off the floor at a time on a Friday after 7pm. Section assignments were rewritten so that the two sections nearest the door were never both covered by someone on the same break wave.
Labour cost: unchanged. Nobody was added, nobody worked longer. This is the appealing kind of fix and also the reason it had been missed — there was no budget line for anyone to notice.
Over the following six weeks, Friday negatives fell from an average of three a week to under one, and the "felt forgotten" phrasing stopped appearing altogether. Total ratings moved from 4.0 to 4.2, which is about what the arithmetic predicts and no more.
Where this reasoning goes wrong
Two honest warnings, because this method produces false positives as readily as findings.
A Friday skew can be a volume artefact. If Friday is 30 per cent of your covers, Friday should be roughly 30 per cent of your complaints. It is only a signal when the share of complaints substantially exceeds the share of covers, and most operators never do that division. Sixty-three per cent against 19 per cent is a finding; 30 against 28 is nothing.
Nine reviews is a small sample that happened to be unusually specific. What made this credible was not the count but the phrase — an unusual, repeated, non-obvious form of words that pointed at one mechanism. A cluster of generic "slow service" reviews on Fridays would not have justified rewriting the break schedule. We have written up a Friday pattern that looked identical and was not real, which is worth reading before you act on your own.
If you want to run it this week
You need three things you already have: your last twenty negative reviews, your rota for the same period, and the break log or an honest account of when breaks actually happen. Group the reviews by day, pull out any stated arrival time, divide the complaint share by the cover share, and then look at what the rota says was happening in the winning hour.
If nothing lines up, stop — the pattern is probably volume. If something lines up as cleanly as it did here, you have found a scheduling problem, and scheduling problems are the cheapest kind to fix.
The reason this does not get done monthly is not difficulty. It is that the hour it takes lands on a Tuesday when two people have called in sick. That is the gap OMMU covers: the daypart read runs continuously and arrives as a short weekly note naming the shift worth looking at, rather than a dashboard waiting for someone to log in.