On this page
Every vendor in hospitality now offers something described as guest experience analytics. Some of it measures things that are genuinely measurable. Some of it produces a sentiment score with a decimal point and an implied precision that the underlying data cannot support.
Since this company has published a study in exactly this area, it is worth setting out where the line sits, including where our own work stops.
What can genuinely be measured
Behaviour in the transaction record
Return frequency, spend per head, product mix, how the mix shifts across an evening, how long a table's sequence of orders takes. This is real, it is already in your point of sale, and almost no bar looks at it in this way.
It is behaviour rather than feeling, which is a limitation and also a strength: nobody has to be surveyed and nothing has to be inferred.
What guests write, at scale
Review text can be classified reliably into themes, and the themes are informative. That is what our study did: 13,063 reviews across 41 Texas venues, classified twice by separate models, then checked against state alcohol tax filings so the classification could be tested against something external rather than only against itself.
Two models rather than one matters. A single classifier's output is difficult to distinguish from its own biases; agreement between two independently built ones is weaker evidence than a controlled experiment and considerably stronger than one model's confidence.
Operational proxies
Wait times, comps, remakes, complaint frequency by shift. These are indirect, and their strength is that they attach to a specific night, which means they can be diagnosed.
What cannot be measured, whatever the dashboard says
Silent churn
The guest who had a poor second visit and simply did not return is invisible to every method available. They did not complain, did not review, and are not identifiable in the transaction record unless you have a loyalty scheme with high coverage. Any product claiming to quantify this is estimating, and the estimate is doing a lot of work.
Sentiment as a precise number
Classifying a review as broadly negative about service is reasonable. Reporting that guest sentiment is 72.4 is a presentational choice rather than a measurement, and the decimal point is there for the same reason a price ends in 99.
Causation from correlation
A venue with more positive reviews and higher revenue has not established which caused which, or whether both follow from something else. Our own study is subject to exactly this limit, which is why it describes relationships and does not claim mechanisms.
Four questions to ask a vendor
What is the underlying data, and is it behaviour or self report. How many reviews or observations sit behind a given score. What is the check on the classification, if any. And what does the product claim it cannot see.
That last one is the most informative question you can ask anybody selling analytics, including us. A vendor with a clear answer has thought about the limits. A vendor without one has not, or would rather you did not.
Where we stop
Our study describes what guests wrote and how it related to trading across 41 venues. It does not tell you what any individual guest felt, how many left without saying anything, or what would have happened had the venue behaved differently. The method is on the research page so you can judge it rather than take our word for it.
What CoreTAP does with the guest side is narrower and more useful day to day: it joins the operational proxies to the specific shift, so a complaint becomes something with a date and a cause rather than a bad feeling about a night.