Somebody will tell you this year that the industry average restaurant rating is 4.3, that good operators respond to 80% of reviews, and that you should be getting a review per sixty covers.
Those numbers are not fabricated. They are also close to unusable, and it is worth being precise about why, because the problem is not that they are inaccurate — it is that they are averages over a population you are not in.
What gets blended together
Any published restaurant benchmark is a mean across a sample. Consider what is inside a typical one:
- Quick-service sites doing 4,000 transactions a week, and tasting menus doing 180 covers
- Delivery-led operations where 70% of orders are off-premise, and pubs where it is 2%
- Businesses open three years with 90 reviews, and businesses open fifteen years with 2,400
- Chains with a central team answering reviews within an hour, and single sites where the owner is also the chef
- Markets where leaving reviews is routine, and markets where it is not
A mean across that is a real number describing nothing in particular. The variance within the sample is larger than almost any difference you could care about, which is the technical version of the complaint.
Three specific distortions do most of the damage.
Star averages are dominated by age. A restaurant with 2,400 reviews cannot move its average; a restaurant with 90 can move it by a tenth in a month. Comparing your 4.1 to an industry 4.3 compares your review count to theirs as much as your food.
Response rate is dominated by structure. Groups with a central reputation function post coverage figures a single site cannot reach without hiring somebody. The published figure is largely a measure of how many respondents were chains.
Review volume per cover is dominated by format and geography. A drive-thru and a destination pub differ by severalfold on this with nothing operationally wrong at either. Delivery platforms prompt for ratings in a way table service does not.
The specific harm of benchmarking against the wrong population
This is not merely imprecise; it produces wrong decisions in two directions.
False alarm. A destination pub with 60% repeat custom and a 4.1 average spends a quarter chasing an industry 4.3 that is structurally unavailable to it — its rating is dominated by six years of history and its guests are not prolific reviewers. Real effort, no movement, demoralised team.
False comfort, which is worse. A delivery-led site sitting at 4.4 against the same 4.3 concludes it is fine, while its own trailing figure has fallen from 4.6 over eight months. It is beating an irrelevant average and losing to itself, and the only number that would have shown this was its own.
The second failure is more common and much more expensive, because nothing prompts anyone to look.
Your own baseline, which takes an afternoon
The alternative is unglamorous and nearly always sufficient. Build a trailing twelve months of your own, on fixed definitions, and compare yourself to yourself.
Step one: export twelve months of reviews. Everything you have access to, with dates, ratings and text.
Step two: bucket by month and calculate four numbers per month. Use the same definitions throughout — the formulas and the denominators they need are worth settling first:
- Reviews received
- Mean rating of that month's reviews — not your cumulative average, which cannot move
- Response coverage, lagged seven days
- Your top three theme rates
Step three: calculate the band, not the average. For each metric, take the twelve monthly values and find the median and the range the middle eight fall in. That band is your normal. A month inside it is a month; a month outside it is an event.
This is the step that converts a baseline into something useful. "Our mean monthly rating is 4.2" tells you nothing on a Monday. "Our monthly rating normally sits between 4.05 and 4.35, and last month was 3.9" is actionable, because you now know that 3.9 is outside twelve months of your own variance.
Step four: mark your own events on the series. The dates of menu changes, price rises, a GM moving, a refurbishment, a new supplier. This is what makes the series diagnostic rather than descriptive — a dip with a menu change two weeks before it is a different conversation from an unexplained one.
Step five: note the seasonal weeks. Some of your variance is annual and repeats. A third of it is usually on the calendar, and comparing August to November tells you about the weather.
When external comparison is legitimate
Two cases, and they are narrow.
Your own direct competitors, read manually. Not an industry average: the four restaurants a guest is actually choosing between. Their recent reviews, their themes, their response behaviour. The sample is tiny and unrepresentative, and that is exactly what makes it relevant — it is the population you compete in. Half an hour reading, once a quarter. The useful output is not their rating, it is which complaints they get that you do not, and vice versa.
Peer groups within your own estate. If you run several sites, a like-for-like internal comparison is the strongest benchmark available anywhere, because the definitions, the market and the brand are held constant. It only works once the peer groups are honestly drawn, and most single-list league tables are not.
The one figure where an external number does help
Response speed, loosely.
Not because an industry median is meaningful, but because guest expectation here is set by platforms rather than by restaurants and does shift over time. If replies broadly arrive within a day or two across hospitality, a venue taking three weeks is visibly slower in a way a guest notices, independent of anything about your own history.
Even then, treat it as a direction rather than a target. The venues posting the fastest medians are the ones with somebody whose job it is, and matching them by posting worse replies faster is not an improvement.
The objections worth conceding
A new site has no baseline. Twelve months of history is the one thing you cannot create on demand. For the first year, external figures are genuinely the only comparison available — use them, loosely, and start building the series on day one so that year two is better.
Self-comparison can be complacent. "We are in line with our own trailing twelve months" is consistent with having been mediocre for a year. The baseline tells you whether you are changing; it says nothing about whether the level is good. The answer to that is reading what your guests actually write, not any number here.
Bands need enough volume. With 15 reviews a month a monthly mean rating swings widely for no reason, and a band built from it will be too wide to flag anything. Use quarters rather than months below about 25 reviews a month.
Twelve months may contain an anomaly. A refurbishment, a flood, three months with no GM. Mark it and exclude it from the band rather than pretending it is normal variance.
You will still be asked for the industry figure. Boards, franchisors and investors ask, and "that number isn't comparable" is a poor answer on its own. Give them the external figure and your own series with your events marked, and let the second one carry the argument. The series is nearly always the more impressive document, because it shows cause and effect, and an industry average never can.
OMMU holds the trailing series with your own events marked on it, which is mostly a matter of not having to rebuild it every time somebody asks. A spreadsheet and an afternoon gets you the same thing — and it is worth the afternoon, because the alternative is managing against the average of a few thousand businesses that are not yours.