The only causal study of restaurant reviews found no effect for chains

Reputation advice in this category is almost entirely assertion. The one piece of real causal evidence is a regression discontinuity study of Seattle restaurants between 2003 and 2009, published in 2011 and revised in 2016, which exploited Yelp's rounding of ratings to isolate the effect of the displayed star from the underlying quality it reflects. It found a 5 to 9% revenue increase per additional star for independent restaurants, and no statistically significant effect for chains. The study is old, it is one city, and it is one platform. It is also the only thing of its kind anybody has, and the asymmetry it found is more useful than the headline number.

The star moved revenue for independents and did nothing measurable for chains, which tells you exactly what a rating is substituting for.

1. 01

   The independents-only result explains what a rating actually does

   A chain arrives with a brand prior: the diner already knows roughly what the food will be and what it will cost, so a rating adds little information. An independent has no prior, and the rating is doing the entire job of predicting the meal. The star is not a quality signal so much as a substitute for one, which is why it moves revenue precisely where no substitute exists.

2. 02

   Cite the study with its age or do not cite it

   2003 to 2009 in Seattle, on Yelp, before smartphones were universal and before delivery apps existed. Anyone presenting the 5 to 9% figure as a current national benchmark is overselling it. Presented honestly it is still the strongest evidence in the category, because the alternative is vendor correlation studies with no identification strategy at all.

3. 03

   74% only consider reviews from the last three months, which makes velocity a running cost

   A rating built two years ago is not doing work today. That converts reputation from a project with an end date into a standing operational motion with a monthly cost, and it is the single most under-budgeted line in local marketing.

4. 04

   47% avoid a business with fewer than 20 reviews, which is a cold-start problem

   A new location opens below the threshold at which roughly half of consumers will consider it at all, on a surface where 68% also require four stars or better. The first twenty reviews are therefore not a marketing task; they are a condition of the location being viable in local search, and they should be scheduled before opening rather than after.

5. 05

   Three numbers this category badly wants do not exist

   There is no credible published figure for restaurant cost per cover, no credible published restaurant customer acquisition cost, and no verified third-party delivery commission percentage. Every widely circulated version traces to a vendor blog or an operator anecdote. The model does not use them, and a plan that depends on one is unfunded rather than optimistic.

A one-star Yelp increase raises revenue **5–9% for independents** and has **no statistically significant effect for chains**.

The model treats review velocity as an operating motion rather than a campaign. The request is attached to the transaction — the receipt, the pickup handoff, the table close — because a request made hours later is a request made to somebody who has moved on. Response has a written standard: who replies, inside what window, and in what shape, with a narrow test for when a review is worth disputing rather than absorbing. Reporting is recency-weighted and per location, never averaged across a group, since the published consumer behaviour is per location and the causal evidence applies to independents specifically.

1. 1

   Weeks 1–2

   Baseline

   Rating, review count and review recency per location and per platform, never averaged

   Owner: OmniFlow

2. 2

   Weeks 2–3

   Response standard

   Owner, response window, tone, and a written test for when a review is disputed

   Owner: OmniFlow + client

3. 3

   Weeks 3–6

   Request motion

   Request attached to the transaction across dine-in, pickup and delivery handoff

   Owner: OmniFlow + front of house

4. 4

   Week 6 onward

   Recency reporting

   Rolling three-month review volume and rating reported per location

   Owner: OmniFlow

5. 5

   Month 6

   Review

   Rating trajectory, review recency, and threshold status against the 4-star and 20-review floors

   Owner: OmniFlow

The model reports rolling three-month review volume and rating per location, on the basis that 74% of consumers disregard anything older. The causal evidence available says the payoff concentrates in independent operators, so the model is written for them and explicitly does not promise the same return to a group trading on brand recognition. No cost per cover and no acquisition cost is modelled, because no credible published figure for either exists in this category and estimating one would be the most dishonest thing on the page. Rating is reported per platform as well as per location, since a strong figure on one surface does not compensate for a weak one on the surface a diner actually opens.

Modelled. Inputs: Michael Luca, Harvard Business School Working Paper 12-016, Reviews, Reputation, and Revenue (2011, rev. 2016); BrightLocal, Local Consumer Review Survey 2026 (1,002 US consumers). Modelled outputs are not a forecast or a guarantee of results. Cited: Michael Luca, *Reviews, Reputation, and Revenue: The Case of Yelp.com*, HBS Working Paper 12-016 (2011, rev. 2016). *(Note the vintage if presented as current.)*.

Every figure below is published research, not a client result. They are the inputs to the arithmetic above, listed so it can be checked rather than taken on trust. The bracketed number points to the full citation at the end of this page.

These are the surfaces this engagement is run and measured from, shown with representative figures built around the benchmarks cited on this page. Every account we run reports into views like these, and you keep ownership of all of them.

Each figure on this page, the system it is read from, and the definition and window it is measured over.

Not applicable — this is a modelled engagement. Its weakest input is the age and narrowness of the only causal study available: one city, one platform, a window ending in 2009, in a market with no delivery apps and no universal smartphones. The consumer threshold figures are current but measure stated intention rather than observed behaviour, and stated intention about reviews is systematically stricter than what people actually do. The model uses the direction of both and instruments the operator's own rating trajectory rather than promising a percentage.

A modelled capability, not a client account. Figures illustrate what the model produces and are labelled as modelled wherever they appear. See all our work

Full citations for everything cited on this page, with the sample and period each study covers, so you can go and read the original.