## Introduction

Worldwide, biodiversity is declining at unprecedented rates, threatening species persistence as well as the benefits humans gain from ecosystems1,2,3. These benefits, known as ecosystem services, have become an increasingly important argument for biodiversity conservation4,5,6,7,8. The economic and other benefits from ecosystems can motivate conservation action, and are more and more being used in payment for ecosystem service schemes. Once an economic value of the service has been determined, it can be captured in commercial markets or quantified in terms comparable with economic services and manufactured capital9. These economic values can then potentially be used to support biodiversity conservation within policies.

The use of ecosystem services arguments for justifying biodiversity conservation is, however, not without risk or controversy. Many experimental studies show that biodiversity increases the magnitude and/or stability of ecosystem functioning (of which ecosystem services are the subset that benefit people), and that most species contribute to ecosystem functioning in some way10,11,12,13. However, such studies do not consider the costs of maintaining or promoting biodiversity, even though costs are generally a limiting factor for implementing real-world conservation policies14. When the economic pay-off from ecosystem services is the main factor motivating conservation, the cost-effective action is to conserve the subset of species that provide the greatest return at relatively short timescales. Because real-world communities are almost invariably dominated by a small number of species15,16 that often respond readily to conservation management17, we hypothesize that in real-world landscapes (1) the majority of the services is provided by a relatively small number of species; (2) that these species are generally common, and that threatened species rarely contribute to present ecosystem service delivery; and (3) that the most important ecosystem-service-providing species can be easily enhanced by simple management actions that are insufficient to support threatened species. Support for these hypotheses would suggest that delivery of ecosystem services is insufficient as a general argument for biodiversity conservation18,19,20,21.

### Identifying dominant crop-visiting bee species

Bee species were characterized as being dominant within a study when their relative abundance on crop flowers was 5% or higher. This threshold corresponds to the cumulative set of species that collectively provide 80% of the crop flower visits (Supplementary Fig. 2). Sensitivity analysis on this choice of threshold showed that results were robust to the choice of threshold so long as the definition of ‘dominant’ did not fall below including species that contributed only 2% of total crop flower visits (Supplementary Fig. 3). Furthermore, our results regarding the dominant crop-visiting species were robust to various study designs and methodological differences among studies, including the spatial extent of sampling and sampling effort (Supplementary Fig. 4). Last, as is often the case for studies of bees for which identification keys do not exist for many parts of the world, there were some unidentified specimens in our studies. These difficult-to-identify taxa were generally rare, however (when pooled, still <5% of the specimens in a given data set), and thus would have minimal impact on our main analyses.

### Crop-visiting bee species relative to regional species pool

Conservation policy objectives are often formulated at national or even continental levels. We therefore also explored how the number of bee species encountered in our studies compared with the total number of unique bee species existing in the political territories in which the studies were performed (that is, the regional species pool). We used a database compiled from published and unpublished sources by J.S.A. of all described bee species currently known to exist in each country, state or province (that is, at the lowest territorial level for which such lists could be obtained). We obtained these data for the German federal states of Hessen55, Lower Saxony56 and Bavaria57, and for the European countries of France, Great Britain, Hungary, Israel, Italy, Netherlands and Sweden (from ref. 58). In North America, species lists were obtained from ref. 58, for the US states California (CA), Massachusetts, New Jersey (NJ), New York, Pennsylvania and Virginia, and the Canadian province of British Columbia. Elsewhere in the world, species lists were used from ref. 58 for Chiapas (Mexico), Costa Rica, Minas Gerais (Brazil), New Zealand, South Africa and Sulawesi (Indonesia). We subsequently calculated straight-forward sample-based species accumulation curves using EstimateS software59, treating each territorial species list as a sample. Because each species list is not an ecological sample but is based on collections, revisions, faunal surveys and national inventories, we refrained from calculating a true species richness estimator.

To examine what proportion of the regional bee species pool visited crop flowers, and what proportion of them was dominant in at least one study, we similarly generated species accumulation curves for (dominant) crop-visiting bee species. Using the full data set of all observed bee species on crop flowers in our data set, we computed the nonparametric, asymptotic true species richness estimator Chao1 with log-linear 95% confidence intervals60, which corrects for unseen species based on the number of species in each study that were observed once (singletons) or twice (doubletons). For dominant species, which included no singletons or doubletons, and further are unlikely to include missing species, we calculated straight-forward species accumulation curves.

### The contribution of threatened species to crop visitation

To examine what proportion of the bee communities observed on crops had a recognized threat status, we used Red Data Books. Red Data Books were only available for four of the countries from which we had data of crop-visiting bee species: Germany61, Netherlands62, Sweden63 and United Kingdom64. In total, 19 separate studies had been carried out in these countries for which we calculated the per study mean pooled proportion of individuals from threatened species.

### Data sets to study commonness and effects of conservation

To address the hypotheses that dominant crop-visiting bee species are generally common species and that these species can be easily enhanced by simple management actions, we used data from a number of European and North American studies examining the effects of measures to promote biodiversity in agricultural areas. These studies used paired designs and standardized protocols to compare bee community composition on sites with biodiversity-enhancing management with that on control sites (sites that were as similar as possible to the treatment sites but were not exposed to biodiversity management). Full details of the study locations and methodologies of the European studies collected in the EU-funded EASY project are given in refs 17, 65. In summary, these sites were sampled in Germany, Hungary, Switzerland, the Netherlands and the United Kingdom in 2003. In each country, three regions were selected with contrasting landscape structure with each region containing seven field pairs. Biodiversity-enhancing management involved delaying the first seasonal cut of grasslands, restricting agro-chemical usage, and/or restricting cattle stocking rates (Hungary, Switzerland and The Netherlands), organic arable farming (Germany) and establishing 6-m-wide grass field margin strips along arable fields (the United Kingdom); all interventions were in the framework of existing agri-environment schemes. In each field, all samples were taken along two 95-m-long transects: one along the field edge and another, parallel to the first one, 50 m from the edge in the grassland interior. We sampled bees using sweep nets (60 sweeps per transect per round) and transect surveys (15 min sampling per transect per round) in the edge and interior of the fields three times (May, June and July) in 2003. For analyses, all data per field were pooled.

In the United States, unpublished 2012 data were used from two studies in CA, one in NJ and one in Michigan (MI). Biodiversity-enhancing management involved establishment of hedgerows of native perennial plants (study CA1) or establishment of wildflower plantings (studies CA2, NJ, MI). In contrast to the European studies, experimental sites in the United States were generally located adjacent to agricultural fields on pre-existing field edges or old fields. For the CA1 study, 20 field edges were selected containing native plant restorations (all at least 5 years old), which were paired with 20 non-restored control sites. Restorations were 350 m long and 3–6 m wide and contained a mix of native perennial shrubs and trees24. Control sites were selected to roughly match conditions surrounding paired restoration sites; for each restoration site, a control site was selected adjacent to the same crop type (row crop, orchard, pasture or vineyard) within the same landscape context (that is, within 1–3 km of the restoration site), but at least 1 km from all other study sites. Control sites were generally weedy field edges and they reflected a variety of unmanaged crop field edges found in the region. Bee communities were sampled at each restoration and control site four times (except one pair of sites sampled only three times). Bees were netted along a 350-m transect for 1 h, stopping the timer while handling specimens. All native bees were collected and identified in the laboratory. The other three studies (CA2, NJ and MI) used the same general approach; each had six site pairs consisting of a wildflower plot established at least 2 years before sampling, using diverse (at least 10 species) mixes of native wildflowers that provided resources for bees throughout the growing season, paired with a control plot that was unrestored. Sampling sites within each pair were separated by 100–800 m. In NJ, four 40 m transects were established within each plot and sampled once in the morning and once in the afternoon, for 10 min each (net sampling time). In MI and CA2, eight 23-m-long transects were established in each plot and were sampled once in the morning and once in the afternoon for 5 min. All bees visiting flowers within 1 m of the transect were collected. In all three studies, each site was sampled four times throughout the summer. Again, for analyses, all data per site were pooled.

### Analysing commonness in relation to semi-natural habitat

To examine whether dominant crop-visiting bee species are common species in agricultural landscapes, generally (hypothesis 2) only data from the control sites were used because they were situated in agricultural habitats such as arable fields (but not flowering, bee-pollinated crops), grasslands, old fields and hedgerows. The proportion of the bee communities consisting of individuals from bee species dominating crop vistitation rates (Supplementary Table 3) were then calculated. The units of analysis were averages of multiple fields, as sample size per site was too low to yield reliable estimates of the relative contribution of dominant species to the bee community. In Europe, averages per region within each country (n=7) were used, whereas in the United States the average per study was used. For the studies MI, NJ and CA2, sample size was six, whereas for CA1 sample size was nine, since land cover data (see below) for all 20 site pairs were not available. To explain differences in the proportional contribution of dominant species between studies, this variable was tested against a number of variables known to affect bee species community composition: the percentage of semi-natural habitat in the vicinity of sampling sites, latitude and continent26. The percentage of semi-natural habitat (for example, extensive grasslands, forests, heathlands and wetlands) was calculated in a radius of 1,000 m around each site, an approximate mean range at which different species groups of bees have been shown to respond to semi-natural habitat in studies on different continents48,66. For the European sites, we used CORINE Land Cover 2006 data sets67 (all land use classes with codes starting with 3 or 4) which, although less accurate than national data sets, provide spatially consistent land cover classifications across all countries. In NJ, land cover data sets provided by the State Department of Environmental Protection were used (http://www.nj.gov/dep/gis/lulc07cshp.html). In MI, land cover was manually digitized from 2012 National Agriculture Imagery Program orthoimagery at the 1:2,000 scale (United States Department of Agriculture Geospatial Data Gateway, http://datagateway.nrcs.usda.gov/). The other two US studies used the National Agricultural Statistics Service crop data file (http://nassgeodata.gmu.edu/CropScape/).

We used standard multiple linear regression models to relate the proportion of individuals from dominant crop-visiting species in bee communities to the proportion of semi-natural habitat, thereby correcting for latitude and continent. Plotting residuals versus fitted values confirmed that model assumptions were met satisfactorily. The often used arcsine transformation of proportional data or binomial regression increased heteroscedasticity, and we therefore present the results of untransformed data. To subsequently explain the patterns in the proportional data, we calculated standardized abundances of dominant crop-visiting bees and, separately, for all other bees for each of the European study regions by dividing the per region bee abundance by the mean abundance across all 15 regions. Since the study in each region had used exactly the same survey protocol, a standardized bee abundance >1 indicates above-average bee abundance compared with the cross-study mean, and a value <1 indicates a below-average bee abundance. We similarly calculated standardized abundances of dominant crop-visiting bees and, separately, all other bees for the three US studies that used the same survey protocol (study CA1 used a different survey protocol and was excluded from this particular analysis). The same approach was used to calculate per study standardized species richness. This allowed us to use the European and US data sets in a joint analysis. We used log-linear models assuming a Poisson distribution with standardized abundance or species richness as response variables, and the proportion semi-natural habitat, bee type (dominant crop-visiting bees versus all other bees) and their interaction as main explanatory variables of interest. A significant interaction would indicate that dominant crop-visiting bees and all other bees are differently related to semi-natural habitat. Latitude was again included as a correcting variable. Continent was not included because we had standardized the response variables between the studies on each continent.

### Analysing effects of measures mitigating biodiversity loss

We used site-level count data as the statistical unit and used generalized linear mixed models assuming Poisson error distribution and using a log-link function68. The initial models used treatment pair as a random term and study, mitigation measure (yes and no) and their interaction as fixed terms. This revealed a significant interaction between the effects of mitigation measures and study (F8,267=3.94, P<0.001). We therefore chose to perform separate analyses for each study with treatment pair as a random factor and mitigation measure as a fixed factor. We chose not to correct for multiple testing, as correction reduces type I error, but tends to inflate type II error69. Instead, we critically interpret statistical outcomes of analyses comparing treatment means. Model outcomes were checked by plotting residuals versus fitted values, confirming that assumptions were met satisfactorily.

All models were fitted using standard facilities in Genstat70.