Title: We altered the title to include non‐alcoholic drinks as well as food because relevant studies rarely evaluated food and drink separately. As a consequence, it was not possible to isolate the effect of labelling for these different products and to restrict the inclusion criteria to food only would have let to the exclusion of several potentially relevant studies. We also altered our inclusion criteria to reflect this change.

Background: we made small changes to the Background to bring the review up‐to‐date, including the addition of more recent references, notably Rayner 2013.

Types of studies: the protocol and review state, "Based on Cochrane recommendations, [ITS] studies that reported only a simple pre and post‐intervention comparisons were not included in the review analysis unless a valid justification for their inclusion could be made or a re‐analysis of the data could enable data from multiple observations in the pre and post periods to be analysed using repeated measures methods" (Cochrane Public Health Review Group 2010; EPOC 2015). We added the following sentence to clarify that studies that either presented appropriate data in graphs, or did not present data in graphs, but did present other types of statistical tests (i.e. other than t‐tests) could be eligible for inclusion: "Authors had to present these data within a graph and/or at the very least analyse them using regression analysis, preferably using segmented regression."

Types of interventions: the protocol stated that a label could be compared to a group in which participants see the same food product presented without a label or with an incomplete label. As we found a number of papers in the search that compared two or more types of labels, we added the following text for clarity: "As noted above, the intervention labelling group had to be compared with a no‐labelling (or incomplete) control group. Thus, we excluded studies that only compared two or more different types of labelling schemes without a control group."

Primary outcomes: the protocol specified purchasing or consumption of foods only, but we also included studies that evaluated the effect of labelling on purchasing of non‐alcoholic drinks (for the reasons stated above). We also added the following sentence to this section of the review for clarity: "We excluded studies that evaluated intention to purchase or intention to consume without an objectively assessed measure of the behaviour."

We also clarified that purchasing had to involve payment with money, as we found some studies in the evidence base that evaluated choices in the settings of interest (e.g. grocery stores), but did not involve purchasing per se.

Food consumption: the protocol specified that where the food consumed was heterogeneous (e.g. a meal comprising various elements with different nutritional content), the amount of each separate element consumed within the meal needed to be assessed for the study to be included. However, this approach would have excluded a number of otherwise good‐quality studies. Thus, we ended up including studies that evaluated multiple food elements and consumption by weighing the meal before and after consumption.

Search methods for identification of studies: we also searched for trials in progress, which was an additional source not specified in the study protocol.

Selection of studies: the protocol stated that "[w]here studies are excluded only on the basis of an incomplete label, the details of these studies will be tabulated separately." We did not identify any incomplete labels, so there was no need to tabulate any details separately.

Data extraction and management: we planned to extract data on any measures relating to the process of implementing the intervention, including data on cost of implementing the intervention in any of the included studies. We did not do this in the final review due to lack of data.

Assessment of risk of bias in included studies:

We added detection bias to this section, which we had omitted. We also removed the risk of bias domain of outcome measurement assessment, as objective outcome measurement was an inclusion criterion for the review. Moreover, we added information to the review regarding how we determined an overall assessment of risk of bias for each study (which we had not specified in the protocol).

The protocol specified that we would use the Quality Assessment Tool for Quantitative Studies (EPHPP 2009) to estimate the risk of bias in controlled before‐and‐after studies and to compare the risk between different types of studies. We did not end up using this tool because we did not identify any eligible controlled before‐and‐after studies.

In addition to the quality assessment strategies specified in the protocol, we conducted a GRADE assessment of the evidence for each outcome according to Cochrane guidance.

Measures of treatment effect: we added the following two paragraphs to the review, which we did not present in the protocol. We added the first paragraph to help quantify the results, and the second because many of the ITS studies were of poor quality, so we considered re‐analysis of the data to be of limited value:

"In order to re‐express effect sizes using a more familiar metric, we calculated the percentage reduction in energy consumed over a typical meal, using an average of 600 kcal as a baseline. This amount was based on mean daily energy intake across the UK population of 1727 kcal or 7226 kJ (standard deviation (SD) 537 kcal or 2247 kJ, using data from the UK National Diet and Nutrition Survey (National Centre for Social Research 2012). Our approach to re‐expressing effect sizes was based on Hollands 2015.

"For ITS studies, we aimed to present statistical comparisons of time trends before and after the intervention (EPOC 2015). In all of the ITS studies, we present the results as described by the study authors, typically as regression analyses. When studies also presented data graphically, we did not attempt any re‐analysis using segmented time series regression techniques if the data were already appropriately analysed by the study authors or if we did not consider the study to be of sufficient quality to warrant re‐analysis. We considered one ITS study to be at low risk of bias (Bollinger 2011), but we could not re‐analyse the data presented graphically due to a lack of information. The figures presented weekly calories per transaction, but there were no data on the number of transactions per week; this means that the absolute and relative variability of each point was unknown and could not be modelled with time series to provide unbiased estimates."

Unit of analysis issues: we updated this section to reflect current methodology. This section now states, "For eligible cluster‐randomised trials, we planned to adjust the data to account for clustering if the study authors had not already done so. However, we only included one cluster‐RCT in the review, and the appropriate data needed to report and adjust the results were not available."

Assessment of heterogeneity: we added a sentence to the review regarding how we would deal with non‐statistical heterogeneity as well as statistical heterogeneity (the protocol only described the latter).

Assessment of reporting biases: the protocol stated that we would use funnel plots to assess reporting biases. However, we could not do this because none of the meta‐analyses included more than 10 studies.

Data synthesis: the protocol stated, "[W]e will only include studies considered to be at lower risk of bias in the meta‐analysis". We included all available evidence in one meta‐analysis but also did a separate sensitivity analysis for studies considered to be at lower risk of bias. We did this in order to present a comprehensive overview of all of the evidence and because we considered very few studies to be at low risk of bias.

A number of unanticipated data synthesis challenges emerged once we identified the included papers. In order to describe how we dealt with these, we added the following text to the review:

1. "For included studies with more than one eligible intervention arm, we combined data when studies contained information about the same product characteristic (e.g. energy), albeit in multiple ways (e.g. varying in whether presented as numbers, colour coded, activity‐equivalents, and whether presented with recommended daily energy intake).

2. "Where studies assessed the impact of nutritional labelling adjacent to a range of food products and it was not possible to extract an effect summary for the range of food products, we included the data for the product representing the most complete meal, for example, sales of entrées (as opposed to sales of a side dish) (e.g. Dubbert 1984). If no products represented more or less complete meals, we extracted data for products containing the greatest amount of energy.

3. "Where studies reported a number of outcomes, such as consumption of a range of different nutrients, we used the most frequently reported outcome among the included studies (e.g. Harnack 2008a). Had outcomes been reported in the same study that related to both increased consumption of healthier foods and decreased consumption of less healthy foods, we would have prioritised the latter."

In addition, after examination of the included studies, we decided to conduct separate analyses for laboratory studies that offered multiple and single food options (which we did not specify in the protocol). We added the following text to the section on data synthesis to describe our rationale: "In the process of conducting the review, it became apparent that the studies also differed in terms of how many labelled options participants had to choose from and what kind(s) of nutritional content the labels described. Participants had to make absolute judgments when given only one labelled option and relative judgments when provided with a myriad of options labelled differently. Thus, we analysed these studies separately."

The exploration of effect modifiers: the protocol specified exploration of 10 possible effect moderators of nutritional labelling using subgroup analysis.

1. Body weight: overweight (> BMI 25 kg/m²) or not overweight (< BMI 25 kg/m²).

2. Dietary restraint in individuals intending to diet: restrained eater or unrestrained eater.

3. Gender: male or female.

4. Label amount formats: relative amounts or absolute amounts of the nutrient or energy.

5. Label signposting: signposting present or absent.

6. The national context in which food was purchased or consumed. Initial examination of the literature indicated that a large proportion of the current research originates in the USA. Thus we compared the effects of nutritional labelling in the USA versus other countries. If there were sufficient variation in the country of study, we would make comparisons between countries.

7. Socioeconomic status: more socially deprived or less socially deprived.

8. Expectations of the taste the food: tastes bad or tastes good.

9. Price of the food: more expensive or less expensive.

10. Immediate context in which food is: purchased in a fast food restaurant or non‐fast food restaurant; or consumed in a real‐world or laboratory setting.

There were sufficient data to analyse only two of these effects (dietary restraint and country). Further, the protocol described the procedure for analysing moderating effects for both continuous and dichotomous outcomes. Given that all data included in the meta‐analysis were continuous, we removed the information about the analysis of dichotomous outcomes and added information about the analysis of the continuous outcomes.

Assessment of heterogeneity: the protocol considers three potential sources of heterogeneity for exploration in subgroup analysis.

1. The impact of the positioning of the label, comparing those that appear on the food package with those appearing in another location, such as on a supermarket shelf.

2. The impact of the information given on the label. First, we planned to compare labels giving information about a range of nutrients versus those giving information about one nutrient. Second, as labels most frequently give energy information, we planned to compare the impact of labels giving information about energy content with labels giving information about other nutrients.

3. The impact of the definitions of healthy purchasing and healthy consumption used in this review. More healthy purchasing is considered to be decreased purchasing of less healthy foods or increased purchasing of more healthy foods, but these may be two separate behaviours. We planned to use subgroup analysis to identify whether they were separate behaviours and this a source of heterogeneity. Similarly, we planned to investigate possible heterogeneity as a consequence of defining more healthy consumption as either decreased consumption of less healthy foods or increased consumption of healthier foods.

Due to lack of information (e.g. many studies did not report on the positioning of the label) and/or lack of differences in label format between the studies, we did not conduct these planned subgroup analyses. Also, given that there were only four studies at low risk of bias, various further subgroup analyses were not possible.

Sensitivity analysis: the protocol stated, "Sensitivity analyses will be conducted to explore the impact of missing data comparing results from available‐case and ITT analysis. Sensitivity analyses will also be used to assess the effects of nutritional labelling on behaviour across studies at both high and low risk of bias, specifically the meta‐analyses will be re‐run including all studies regardless of their risk of bias. Additionally, the impact of the definition of nutritional labels used in this review will be explored. The meta‐analyses will be re‐run including the studies excluded from the main analyses due to the presentation of an incomplete label rather than a complete label (as described in the 'Description of the intervention')."

We did not conduct these analyses because only four studies at low risk of bias were available for analysis, which is not enough to enable comparison in the above variables.