Long‐term experimental drought alters floral scent and pollinator visits in a Mediterranean plant community despite overall limited impacts on plant phenotype and reproduction

Pollinators are declining globally, with climate change implicated as an important driver. Climate change can induce phenological shifts and reduce floral resources for pollinators, but little is known about its effects on floral attractiveness and how this might cascade to affect pollinators, pollination functions and plant fitness. We used an in situ long‐term drought experiment to investigate multiple impacts of reduced precipitation in a natural Mediterranean shrubland, a habitat where climate change is predicted to increase the frequency and intensity of droughts. Focusing on three insect‐pollinated plant species that provide abundant rewards and support a diversity of pollinators (Cistus albidus, Salvia rosmarinus and Thymus vulgaris), we investigated the effects of drought on a suite of floral traits including nectar production and floral scent. We also measured the impact of reduced rainfall on pollinator visits, fruit set and germination in S. rosmarinus and C. albidus. Drought altered floral emissions of all three plant species qualitatively, and reduced nectar production in T. vulgaris only. Apis mellifera and Bombus gr. terrestris visited more flowers in control plots than drought plots, while small wild bees visited more flowers in drought plots than control plots. Pollinator species richness did not differ significantly between treatments. Fruit set and seed set in S. rosmarinus and C. albidus did not differ significantly between control and drought plots, but seeds from drought plots had slower germination for S. rosmarinus and marginally lower germination success in C. albidus. Synthesis. Overall, we found limited but consistent impacts of a moderate experimental drought on floral phenotype, plant reproduction and pollinator visits. Increased aridity under climate change is predicted to be stronger than the level assessed in the present study. Drought impacts will likely be stronger and this could profoundly affect the structure and functioning of plant–pollinator networks in Mediterranean ecosystems.

There is an urgent need to document the causes and mechanisms of pollinator decline and especially the role of climate change to predict and prevent further disruption of pollination functions.
Climate change may negatively affect plant-pollinator interactions in many ways , by causing phenological mismatches between the period of activity of pollinator species and the flowering period of their host plants (Duchenne et al., 2020;Memmott et al., 2007), shifts of their geographical range (Rasmont et al., 2015), and a change in floral resources (pollen and nectar; Waser & Price, 2016) which compose the bulk of pollinators' diet. Indeed, nectar production is sensitive to the quantity and seasonality of rainfall, and production is generally lower in drier conditions ( Kuppler & Kotowska, 2021) especially in temperate regions (Gallagher & Campbell, 2017;Phillips et al., 2018;Waser & Price, 2016). Nectar production is also sensitive to temperature and many species reach maximum production at an optimal temperature (Jakobsen & Kritjansson, 1994;Petanidou & Smets, 1996). Droughtinduced reduction in the production and timing of floral resources available to pollinators could therefore lead to declines in their abundance (Forister et al., 2018;Thomson, 2016;Timberlake et al., 2020).
Although much less studied, climate change can also disrupt plant-pollinator interactions in a more subtle way, by altering floral signalling and attractiveness (Walter, 2020). Pollinators use and learn visual and olfactory flower signals-including floral display (number of flowers), flower size and colour, and floral scent-to detect, recognize, and locate their preferred floral resources (Burkle & Runyon, 2019;Giurfa & Sandoz, 2012;Raguso, 2008). Floral scent is a blend of many different volatile organic compounds (VOCs) emitted by floral tissue that guide pollinators towards floral resources. The activation of plant defences also leads to the emission of specific VOCs from inflorescence tissue, which blend in pollinator-attractive floral scent plumes (Borghi et al., 2019;Cunningham, 2012;Wright & Schiestl, 2009).
Therefore, floral scent is highly sensitive to increased temperature (Farré-Armengol et al., 2014;Yuan et al., 2009) and drought, with increased emissions under moderate stress (Burkle & Runyon, 2016;Campbell et al., 2019;Glenny et al., 2018). Altered floral attractiveness could therefore affect pollinators' foraging behaviour and resource choices (Flacher et al., 2017. For instance, herbivoreinduced plant volatiles present in floral scent tend to deter pollinators from floral resources in attacked plants Kessler et al., 2011). Similarly, water-limited plants can be less visited by pollinators due to changes in flower attractiveness and floral rewards (Descamps et al., 2018;Rering et al., 2020).
More broadly, climate change and especially drought affects plant physiology and development. Drought reduces stomatal conductance and therefore carbon uptake, reducing photosynthesis and plant growth (Prieto et al., 2009;Reddy et al., 2004;Saunier et al., 2018), and plant reproduction (Ogaya & Peñuelas, 2007). The emission of plant volatiles following the activation of plant defences is generally increased under moderate drought and reduced under severe or chronic drought (Blanch et al., 2007;Ormeño et al., 2007;Saunier et al., 2017), because of restricted carbon acquisition (Staudt et al., 2002) and down-regulation of carbon-consuming functions such as VOC emissions (Rennenberg et al., 2006). Plant species in semi-arid regions such as the Mediterranean Basin are generally drought adapted and have developed drought escape, tolerance or avoidance resistance mechanisms, notably via increased carbon uptake efficiency (Aslam et al., 2015;Nardini et al., 2014). Yet, under climate change, the drought resistance threshold of Mediterranean plant species and communities may be exceeded, reducing ecosystem productivity and threatening plant diversity (Malone et al., 2016;Rodriguez-Ramirez et al., 2017).
Mediterranean regions within Europe are predicted to be particularly affected by climate change, with a predicted average reduction in rainfall of up to 30% in summer and 10-20% in spring by the end of this century (Giorgi & Lionello, 2008;Mariotti et al., 2015). In addition, this region is a biodiversity hotspot, especially for pollinators (Nieto et al., 2014: Orr et al., 2021, and may be more fragile than some other biomes. For example , Newbold et al. (2020) predicted a disproportionate reduction in species richness in regions with Mediterranean climate, which host more species closer to their temperature limit compared to other biomes. A more thorough evaluation of climate change impacts on plant-pollinator interactions and pollination functions in this region is therefore needed. could profoundly affect the structure and functioning of plant-pollinator networks in Mediterranean ecosystems.

K E Y W O R D S
climate change, floral traits, plant fitness, pollination, reduced rainfall, volatile organic compound, water deficit, water-limited In this study, we investigated the impacts of long-term drought (in situ experimentally reduced rainfall) on plant-pollinator interactions in a natural Mediterranean shrubland community. We compared a suite of floral traits between plants growing under control and amplified drought conditions including the number of flowers, flower size and colour, floral scent, and nectar production. We also quantified pollinator visits to these plants and measured plant fitness. We asked (i) how does drought alter floral traits?; (ii) how do pollinators respond to drought-induced changes in floral traits? and (iii) how do changes in visiting pollinator community coupled with drought affect plant maternal fitness? We expected drought to affect all three aspects, and in particular to cause increased floral scent emissions, reduced nectar production, reduced pollinator visits and reduced fruit and seed production.

| Study site: CLIMED long-term drought experiment
All field data were collected in February-June 2018. We used a subset of established plots that were part of the CLIMED ( Drew et al., 2017), and Cistus albidus Linnaeus, 1753 (Cistaceae; Montès et al., 2008). Local cumulative precipitation between January and May 2018 (the flowering period surveyed) reached 291 mm, while the average precipitation between January and May for the period 2008-2018 was 205 mm (Marseille-Marignane meteorological station; www.infoc limat.fr). The site is equipped with 46 metallic control and 46 4 × 4 m rain-exclusion shelters established in October 2011, spaced by 1-30 m . Plot locations were chosen randomly at the time of establishment of the long-term experiment, and were assigned at random to control or drought treatment (Montès et al., 2008). Gutters from rain-exclusion shelters in drought plots were designed to exclude up to 30% and excluded on average (±SE) 12 ± 2% of precipitation between 2011 and 2018 at the centre of the plots; the intercepted water was carried away from the site with a pipe system. In control plots, the upside-down gutters intercepted a very small fraction of precipitation and rainfall reached the ground (Montès et al., 2008;Santonja et al., 2017). This water deficit attempts to mimic the mean predicted changes during the dry season in the Mediterranean area by the end of this century except in winter when rainfall is expected to increase (Giorgi & Lionello, 2008: averages for 2071-2100relative to 1961-1990 December to February +0 to +10%, March to May −10 to −20%, June to August −20 to −30%, September to November −0 to −10%; Mariotti et al., 2015Mariotti et al., : averages for 2071Mariotti et al., -2098Mariotti et al., relative to 1980Mariotti et al., -2005 December to February −0.1 to +0.2 mm/day, June to August −0.1 to −0.3 mm/day). The moderate but chronic experimental water deficit induced by the CLIMED experiment can alter plant physiology: carbon assimilation was reduced in C. albidus, and transpiration was reduced in C. albidus and S. rosmarinus but water use efficiency was not significantly changed in 2014 (Rodriguez-Ramirez, 2017). Between January and May 2018, permanent soil moisture probes (TDR100, Campbell Scientific Inc., Logan, Utah) measured soil moisture at 10, 20 and 40 cm in two control and two drought plots. For clarity, we use the term drought to refer to the drought treatment in our study.
We selected 10 control plots and 10 drought plots out of the 92 plots, based on (i) where Thymus vulgaris Linnaeus, 1753 (Lamiaceae) was present (four plots for each treatment only) because it is an important resource for pollinators (Ropars, Affre, Aubert, et al., 2020); and (ii) a high and similar percentage cover of C. albidus and S. rosmarinus ( Figure 1). The chosen control and drought plots were homogeneously distributed throughout the site ( Figure S1). We measured the percentage cover of each species in selected plots twice (February and June 2018 on average respectively in the 20 plots selected, and the community composition did not differ significantly between treatments throughout the long-term experiment (Table S4). Despite such low diversity, this plant community is natural, and is representative of the site and of the type of dense, closed vegetation plant communities found in the region in areas where wildfires are ancient (>10 years; Pimont et al., 2018). Thymus vulgaris, C. albidus and S. rosmarinus are all perennial, entomogamous shrub species; T. vulgaris is gynodioecious and obligate entomogamous (dichogamous; Arnan et al., 2014), while S. rosmarinus and C. albidus are self-compatible but with limited self-pollination (Blasco & Mateu, 1995;Hammer & Junghanns, 2020). A fourth shrub species, Ulex parviflorus Pourr., 1788, was also present but very rare (0.3% percentage cover) with very few flowers during the study period, and other flowering species were even rarer (Table S3). We did not observe any insect visit to these very rare species and hence excluded them from our study.

| Floral scent
We randomly selected up to 14 plant individuals per species in each treatment (control vs. drought) with a maximum of two (four for T. vulgaris) plants in the same plot. A few samples were lost during laboratory analysis, hence final sample sizes were 23 (control: 11; drought: 12) for S. rosmarinus, 22 (control: 11; drought: 11) for C. albidus, and 19 (control: 6 female, 6 hermaphroditic; drought: 5 female, 2 hermaphroditic) for T. vulgaris. Floral scent sampling and data processing are described in detail in Supporting Information (1.1).
Briefly, the collection of floral scent volatiles was performed via dynamic headspace sampling in situ. Inflorescences from plant canopy bearing around 30-50 (S. rosmarinus), 2-3 (C. albidus) or 100-400 (T. vulgaris) flowers (1st-3rd quantiles) were enclosed in 2 L bags and placed under a constant flow of purified air (inlet 1000 ml min −1 , outlet 200 ml min −1 ). Two leaf-only scent samples in each plant species and treatment were also collected as a comparison, enclosing branches of comparable sizes than the inflorescences in floral samples (Supporting Information 1.1 and 2.1). VOCs were adsorbed on a Carbotrap/Tenax cartridge (Sigma-Aldrich) placed at the bag outlet for 10 min for S. rosmarinus and T. vulgaris, and 15 min for C. albidus, so as to not exceed the breakthrough volume of each VOC (Ormeño et al., 2007). Ambient air (one sample every three plant scent sample) was also sampled as blank control and temperature recorded for emission rate normalization (see below).
Samples were analysed using gas chromatography (GC) coupled with mass spectrometry after thermodesorption. The methodology is detailed in Supporting Information 1.1. Briefly, peaks were ex- Individual peak mass spectra were compared with that of 21 pure standard molecules injected separately as well as reference spectra (libraries Adams, 2007 andNIST, 2011) using the R function 'SpectrumSimilarity' (R library 'OrgMassSpecR v0.5-3'; Stein & Scott, 1994) and with a tolerance in retention index of ±15 between the analysed peak and the reference molecule. Despite these precautions, the identification remains tentative. Only identified VOCs previously reported as plant volatiles, with an average similarity >0.8 and whose area exceeded three times that of ambient air samples in average were retained for analysis (Campbell et al., 2019 and see Supporting Information 1.1). Emission rates of each VOC (in μg h −1 g DM −1 ) were calculated by subtracting the quantities in their corresponding ambient air sample and normalized by the total dry mass of the bagged inflorescence and by temperature (Ormeño et al., 2007;Sabillón & Cremades, 2001).

| Nectar production
For each plant individual on which we sampled floral scent, we also selected five flowers (only 2.5 on average for C. albidus) to measure nectar production (as well as flower size and colour, Supporting Information 1.4). We measured nectar standing crop (otherwise referred to as nectar production for simplicity) in each flower using 0. (in μg/μl) and total sugar content (in μg) were calculated a posteriori using the conversion table from ° Brix to g/L by Kearns and Inouye (1993) and by applying the dilution factor and a temperature correction after the manufacturer's correction table .

| Plant-pollinator interactions
To quantify plant-pollinator interactions, pollinator visits were ob- Bombus gr. terrestris is by far the most abundant Bombus group species in the region and the other very rare species are functionally similar (Ropars, Affre, Aubert, et al., 2020). No wild Apis mellifera colony was recorded close to our study site and we assumed that 100 % of A. mellifera we observed were managed (see also Herrera, 2020).
For clarity, we refer to Apis mellifera as the managed honeybee in this study. The six pollinator groups defined above are easily identified in situ. Visits by non-Apoidea Hymenoptera and by Lepidoptera were extremely rare and therefore not considered. Observations were carried out under optimal weather conditions, that is, on sunny days with temperatures above 15°C and without strong wind, between 10:00 and 16:00 from 9th March to 5th April and from 8:00 to 14:00 from 13th April to 23rd May. Observations were recorded in the 20 plots (grouped in four blocks) successively in a block-randomized order at each sampling date. Contrary to our expectations and because T. vulgaris was rare (but always selected in the observed most flower-dense patches when flowering), insect visits to T. vulgaris were rare. In particular, all 85 recorded visits by A. mellifera on T.
vulgaris took place in a single plot on a single day. We therefore discarded these visits from our analyses.
Immediately after the 5 min observation round in each plot, visiting insects were caught with a sweep net during the following 5 min, adding 30 s to that duration for each insect caught to account for insect handling time and disturbance to the patch. Insects were then frozen overnight and prepared for identification in the laboratory.

| Plant reproduction
As an estimate of plant female fitness, we measured the fruit set of each marked plant individual of C. albidus and S. rosmarinus (two per plot). As a proxy for fruit set, we counted the total number of fruits per m 2 with the same method as for the total number

| Statistical analyses
All statistics were performed using R version 3.6.3 (R Core Team, 2020).
We performed multivariate analyses of floral scent and pollinator community compositions (detailed below). Otherwise (unless specified), we used linear mixed model ( (Hervé et al., 2018). Typical metabolomic datasets often comprise many more molecules than samples, and these molecules or explanatory variables are often strongly inter-correlated, notably due to shared metabolic pathways . The canonical powered partial least squares discriminant analysis overcomes these biases by combining classification and regression and is commonly used for the analysis of chemical datasets (function 'cppls' with parameters 'centre' and 'scale' to true, R library 'pls'; Mevik et al., 2019;Indahl et al., 2009;Hervé et al., 2018). We implemented the treatment as factor and three com-   Table S12), and such data were also used to estimate sampling completeness (methods in Table S13). Drought impact on the abundance and species richness were analysed using t-tests. We also performed a constrained correspondence analysis to test whether drought affected community composition, using a permutation test with 999 permutations and a Fisher test (functions 'cca' and 'anova.

| Other floral traits
cca', R library 'vegan'; Oksanen et al., 2019) on the same species used for species richness, but removing singletons.

| Plant reproduction
The drought impact on the number of fruits per m 2 was tested with a GLMM with a negative binomial error distribution for each species (Section 2.5.1).
For each species, the impact of drought on the mean number of seeds per fruit, the variance in seed number per fruit (multiplied by ten and rounded), and the mean seed mass was tested using a LMM, a GLMM with a negative binomial error distribution, and a LMM, respectively.

| Efficacy of the rainfall reduction in 2018
Between January and May 2018, soil moisture was lower in drought plots than in control plots by 1.    Figure 2).

Finally, total emissions of oxygenated FADs (including the GLVs)
were significantly greater in drought conditions in S. rosmarinus, but the difference was not significant for the other two species (Table S8). Total emissions of other chemical families did not differ significantly between treatments in any species.

| Nectar production and other floral traits
In T. vulgaris, drought caused a marginally significant 38% reduction in the proportion of flowers producing nectar, and a reduction of 49% in sugar content per flower. No significant effects on nectar production were observed in S. rosmarinus and C. albidus (Table 2; Figure 3).
Total number of flowers was not affected by drought in either of the three species, and neither was flowering phenology (Table S9; Figure S2). Drought did not affect flower size or flower colour significantly in any of the species, although the yellow centre of C albidus flowers was marginally brighter in plants under drought (Tables S10 and S11; Figure S3).

| Plant-pollinator interactions
We recorded a total of 6576 flower visits, including 6064 on S. ros-

| Plant reproduction
Drought did not affect the number of fruits per m 2 for either S. rosmarinus or C. albidus, nor did it affect the mean and variance in seed number per fruit or the mean seed mass (Table 4; Figure 6).
Finally, drought had a limited impact on seeds (Appendix Section 2;  Figure S5D).

| DISCUSS ION
Our study assesses the impacts of an induced long-term experi-

| Limited impact of drought
The overall limited impact of drought on floral traits and plant reproduction was unexpected, since many studies have reported changes in nectar production, flowering phenology, flower size and number, and fruit or seed set under possibly more intense drought (reviewed in Borghi et al., 2019 andDescamps et al., 2021) including in Mediterranean, drought-adapted shrub species (del Cacho et al., 2013). A possible explanation for the lack of response in our system is that Mediterranean plant species have evolved drought resistance mechanisms that could mitigate drought impacts (Aslam et al., 2015;Nardini et al., 2014). These mechanisms or traits may have been further selected for in our experimental drought plots over 6 years, through differential mortality, increasing drought resistance in the plant community (Rodriguez-Ramirez, 2017; ). An alternative, or perhaps complementary explanation is that our study was conducted during a period of unusually high rainfall (42 % higher than the long-term average; Section 2.1), resulting in only modest differences in soil moisture between treatments (Section 3.1). Heavy rainfall events (≥5 mm h −1 )

Rodriguez
could have caused water run-off on the ground into drought plots and causing the especially low difference in soil moisture at shallow depths. The small difference in soil moisture possibly led to high enough water availability to not affect plant performance in drought plots, reducing overall differences in floral traits between experimental and control plots during our study year (Abbaszadeh et al., 2020;Pérez-Llorca et al., 2019).
Only T. vulgaris had a drastically reduced nectar standing crop under drought (total sugar reduced by ~75% when combining sugar per flower and flowers with nectar). This suggests a lower drought resistance than S. rosmarinus and C. albidus. Thymus vulgaris is widely characterized as drought sensitive, compared to congeneric species (Ashrafi et al., 2018) and has also shallower roots than S. rosmarinus and C. albidus. Drought sensitivity however may also have been in- Walter, 2020). Our experiment did not make possible to disentangle these two mechanisms since no hand pollination or pollinator exclusion treatments were included. We did find a limited impact of maternal drought (drought during production of seeds by maternal plants) on seed germination. Similar visit frequencies in control and drought plots could ensure pollination success and seed set production in the generalist plant species (or at least, species visited by a range of pollinators), although changes in the visitor community composition may have led to slight changes in the pollination efficiency (Burkle & Alarcón, 2011). Self-pollination is limited in the species studied (Blasco & Mateu, 1995;Hammer & Junghanns, 2020), but how drought may affect self-pollination rate is unknown.
Alternatively, the minor changes measured in seed germination may be due to a physiological plant response to drought. Slower germination of seeds from drought-stressed S. rosmarinus plants suggests more poorly resourced seeds. It could make them more vulnerable to adverse environmental conditions, which are more likely under climate change (Quintana et al., 2004), so it is difficult to see how this is adaptive. However, the marginally lower germination rate in C.
albidus floral scent under drought here were also shown in C. albidus leaves under drought (Ormeño et al., 2007). In contrast, in S.
rosmarinus, the sesquiterpenes E-Caryophyllene and α-Humulene, emitted at higher rate in floral scent under drought and characterized as typical floral VOCs here (Table S5) were absent in S. rosmarinus leaf samples in a previous study (Ormeño et al., 2007). This may be because these sesquiterpenes have a dual function (defence and pollinator attraction; Schiestl, 2010), and the plant increases investment in bee-attractive molecules (Abraham et al., 2018;Leonhardt et al., 2014) to counterbalance other negative effects of drought (Kuppler & Kotowska, 2021). Drought stress also often increases

| Can altered floral scent explain differences in pollinator visits?
We found that drought altered the relative number of visits by different pollinator functional groups. Workers of A. mellifera and B. gr.
terrestris visited more S. rosmarinus flowers in control than in drought plots, while the species-rich group of small wild bees visited more S.
rosmarinus flowers in drought than control plots. The same trend, although not significant, was found in C. albidus. Other studies have found that a variety of bee species prefer non-water limited plants (Al-Ghzawi et al., 2009;Descamps et al., 2018;Höfer et al., 2021;Rering et al., 2020).
The measured floral scent alterations may partially explain such changes in pollinator visits. Altered emissions could negatively affect flower attractiveness to pollinators and may cause the rewiring of pollination networks, that is, the modification of the relative intensity of plant-pollinator interactions within the community (Larue et al., 2016). The attractiveness of most VOCs to most pollinator species remains unknown, yet among the VOCs altered by drought in our three species, some are described as attractive to A. mellifera and other bee species (Table S14). Conversely, the stress marker 3Z-Hexenol acetate was found to be slightly repellent to B.
terrestris (Ceuppens et al., 2015). The stress-induced increased emission rate of VOCs with a defensive primary function (rather than pollinator attraction) could deter pollinators (Schiestl et al., 2014;Theis, 2006). Avoiding drought-stressed plants based on floral scent signals may be positively selected, or based on associative learning (Jaworski et al., 2015), if altered floral scent is a cue for reduced quality or quantity of floral resources (Wilson Rankin et al., 2020).
While a change in pollinator behaviour is unlikely to be motivated by altered floral scent alone especially in generalist pollination systems, pollinators may use it as evidence for altered floral resources in resource-limited plants . We did not detect changes in nectar standing crop in either S. rosmarinus or C. albidus, but we did not measure flower refill rate, nectar composition (ratio of different sugars and nutrients content), or pollen production and quality, which could have been affected by drought, potentially influencing bee choice (Petanidou, 2005). Wilson Rankin et al. (2020) showed that drought reduced nectar quality and quantity as well as pollen nutritional quality in Trifolium willdenovii, reducing colony fitness in A. mellifera and B. impatiens. The increased visitation rate towards plants experiencing drought by small wild bees (including some Andrena, Osmia and Lasioglossum species, Table S12) is more difficult to interpret, partly because of the diversity of insect species within this group. One speculative hypothesis is that, rather than being a response to floral scent, it is a signal of resource partitioning due to interspecific competition. Apis mellifera dominates the flower visitor community (81% of all visits) in our study and throughout the Mediterranean Basin (Herrera, 2020). A study TA B L E 2 Drought impact on nectar production. Significant effects (p < 0.05) and marginally significant effects (p < 0.07) of drought are shown in bold. with the same observation methods and in a similar habitat 20 km away from our field site in 2017-2018 found that the abundance and species richness of large wild bees was negatively affected by increased honeybee hive density on wild pollinators . The social structure of A. mellifera hives (all managed in the region), and their ability to communicate the location of resources, enables workers to exploit the most rewarding floral resources and to track resource availability much more efficiently than wild pollinator species (Hasenjager et al., 2020;Hung et al., 2019).
The higher number of visits by A. mellifera to control plots suggests that they evaluated floral resources to be more abundant or of higher quality in control plots. If so, resources in the control plots would be more rapidly depleted, leaving resources in drought plots to other pollinator species and causing resource partitioning Thomson & Page, 2020). We observed some A. mellifera  Table 2. * Indicates a significant difference at p < 0.05 ( Table 2).

| Limitations of the study and next steps
Our study is based on just one year of data collection, and this prevents an understanding of how the abnormal rainy conditions might have affected our findings. Replicating the observations over successive years would help disentangle the effects of long-term, experimentally induced drought trends from the effects of shorterterm extreme climatic events. Also, the CLIMED long-term experiment induces drought throughout the year, whereas climate change is predicted to cause more frequent and intense summer droughts, but wetter winters (Giorgi & Lionello, 2008;Mariotti et al., 2015). A more realistic drought simulation (more intense, but shorter) could more strongly affect flower attractiveness and the production of flower resources, with negative consequences for pollinator interactions (Walter, 2018(Walter, , 2020. Another limiting peculiarity of our study is the relatively low plant species richness at the study site (Table S4) Significant differences are highlighted with *p < 0.005 and **p < 0.001; see Table 3.

TA B L E 4
Drought impact on fruit set and seed set. Our experiment was not designed to disentangle the relative contributions of indirect drought, pollinator-induced and direct droughtinduced impacts on plant reproduction, and for this reason the minor changes measured in plant reproduction cannot be unambiguously attributed to either observed changes in the flower visitor community (Kevan & Eisikowitch, 1990) or to plant physiology (Karimmojeni et al., 2014).
Our study was similar to an in situ choice experiment, since foraging bees could freely choose between control and drought plots (Nordström et al., 2017). In reality, climate change is likely to affect entire plant communities and up to a regional scale, leaving no such choice to pollinators. It would be useful-but logistically challenging-to further investigate climate changes impacts at a community scale. For example, this could be attempted by choosing pairs of comparable large-scale communities such as entire valleys in different climatic conditions to assess the impact on pollinator foraging behaviour and population dynamics. Despite climate change, any habitat remains a mosaic of micro-climatic conditions (Maclean, 2020) and our experiment therefore captures some of the realistic climate change predictions. Also, climate change may have direct impacts on pollinator behaviour, phenology and population dynamics (Becher et al., 2014;Burkle et al., 2013;Woodard, 2017), which were not considered here, and which would deserve further attention in integrative approaches.

| CON CLUS IONS
Our study showed that reduced rainfall altered floral scent in the three species studied, reduced nectar production in T. vulgaris only, and caused a shift in the flower visitor community in a Mediterranean system. Drought impacts were otherwise limited on floral traits and rewards, pollinator visits or reproductive success.
Pollinators may adapt to altered floral olfactory and visual signals (Jaworski et al., 2015) but qualitative and quantitative changes in floral resources and therefore potentially in pollinator diet breadth (Schweiger et al., 2010) will affect pollinator fitness and this would deserve further attention. As next steps, we recommend investigating drought impacts in communities where T. vulgaris is more abundant, since this species did show a reduction in floral nectar reward in our experiment, in more diverse Mediterranean plant communities, and in communities with clearly identified specialized and generalist plant and pollinator species. Reduced floral resources and altered pollination functions may result in population declines in both pollinator and plant communities (Wagner, 2020), reducing the effectiveness of pollination functions and ecosystem productivity F I G U R E 6 Drought impact on the number of fruits per m 2 (left), the mean seed number and variance in seed number per fruit (middle), and the mean seed mass (right) in S. rosmarinus (top row) and C. albidus (bottom row). Sample sizes are provided in Table 4.
in biodiversity-rich but also already fragile Mediterranean ecosystems. Under predicted climate change, those ecosystems will also likely endure a combination of extreme events such as intense drought episodes and heat waves of higher frequency and intensity.
This is likely to exacerbate the effects we observed on flower attractiveness, plant-pollinator interactions and plant reproduction.
Quantifying these impacts will be essential to estimate the resilience of Mediterranean ecosystems under ever-increasing anthropogenic pressures.

CO N FLI C T O F I NTE R E S T
All authors have declared to have no conflict of interest.

PEER R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/1365-2745.13974.

DATA AVA I L A B I L I T Y S TAT E M E N T
Data available from the Dryad Digital Repository https://doi. org/10.5061/dryad.h70rx wdmz (Jaworski et al., 2022a). The repository contains all datasets as well as a metafile and readme file. R code used for data analyses available from the Zenodo Repository https://doi.org/10.5281/zenodo.6914377 (Jaworski et al., 2022b).