Most review handling goes wrong at the point of judgement, not the point of writing. The same review gets a defensive reply on Friday and a generous one on Tuesday, because the decision is being made by whoever read it and how their shift went.
A decision tree fixes that by taking the judgement out. Five questions, asked in this order, stopping at the first yes.
Question 1: Does it allege harm?
Illness, allergic reaction, injury, foreign object, a hygiene claim, or anything involving a child's safety.
If yes: do not reply. Escalate immediately to whoever owns incidents, and follow an incident process rather than a review process. A public reply written before the facts are established creates a record you cannot amend, and "we take food safety very seriously" as a first response reads as an admission to everyone who has ever seen it before.
There is a full playbook for this and a first-hour timeline. Neither of them starts with writing.
Stop here. This is the only category where the 24-hour clock is the wrong clock — it is a one-hour clock.
Question 2: Does it name a member of staff negatively?
If yes: get the shift's account before replying, and reply within 24 hours anyway.
The order matters. You need the staff member's version because a public reply that accepts the guest's account uncritically is something your team will remember much longer than the guest will. But you cannot wait three days, because the review is being read now.
The workable form is a reply that acknowledges the experience without ratifying the characterisation: "That's not the standard we hold, and I'd like to understand what happened — could you email me at andrew@ommu.ai?" True regardless of what the investigation finds.
Stop here.
Question 3: Is the factual claim checkable in under five minutes?
A wait time, a charge, a booking, a dish that was allegedly unavailable.
If yes: check it, then reply. This is the largest category of stalled reviews in most venues, and it stalls because the check is a small errand and small errands do not get done. Batch them — one trip to the booking system for all of them, as in the fifteen-minute breakdown.
If it would take longer than five minutes: reply without the fact, and do not assert anything you have not verified. "I don't want to guess at what happened" is a legitimate and rather good sentence.
Stop here.
Question 4: Is it negative, or a four-star with a specific caveat?
If yes: reply within 24 hours.
Negatives are what future guests read. Four-stars with a caveat are the most useful thing you will receive and have their own reply style — do not apologise into them.
Stop here.
Question 5: Is it a plain positive with no detail?
Reply to a rotating sample, in genuinely varied words. Not all of them, and never with the same sentence. Forty identical thank-yous on one profile is worse than twenty replies and twenty gaps, because the repetition is visible at a glance and it undermines the credibility that replying was supposed to buy.
The bucket the tree does not have
There is no "ignore" outcome, deliberately. Almost every triage framework has one, and it is where reviews go to be forgotten — including a few that turn out to matter. Reviews that reach question five without a yes are not ignored, they are sampled, which is a different instruction with a different failure mode.
The one genuine exception is content that is not a review at all: obvious spam, a review for another business, abuse with no substance. Those go to a report process, not a reply process, and the reviews you decided not to answer is about managing that pile properly rather than letting it accumulate.
Making it stick
Print it. The tree only earns its value when the duty manager can run it at 9pm without calling anyone, and that requires it to be on a wall rather than in a document.
Two things to write on it before you post it up: who owns question 1 escalations by name, and the email address used in question 2 replies. A tree with an unnamed owner produces a pause, and a pause on a harm allegation is the expensive kind.
The remaining problem is that the tree has to be run on every review, every day, including the day two people call in sick — which is the part that fails rather than the reasoning. OMMU runs the sort continuously and texts the ones that answer yes to questions 1 or 2, so what needs a person reaches a person and the rest does not require a session.