Field experiments were conducted at two study sites – Vale Carapito (Site 1) and Ermo das Águias (Site 2) – in the Greater Côa Valley region in Portugal ( Figure 1 ). This region has a Mediterranean climate, characterized by hot and dry summers, and cold winters. Both sites are characterized by woodland-shrubland mosaic habitats, with dense shrub patches dominated by Spanish broom (Cytisus multiflorus) and woodlands dominated by holm oak (Quercus rotundifolia) and Pyrenean oak (Quercus pyrenaica). Site 1 and Site 2 were grazed by small ruminants (sheep and goats) for the 40–60 years prior to being grazed by cattle in 2010 and 2005, respectively ( Supplementary Table S1 ). The sites were then acquisitioned by Rewilding Portugal in 2020 and 2021, respectively. Site 1 experiences a wetter climate, with an average annual temperature of 12.7°C and precipitation of 795 mm ( Supplementary Figure S1 ), and encompasses 65 hectares of grazing area. The grazing regime at Site 1 changed in 2021 from a baseline regime of 25 cattle under rotational grazing (2–3 months, two times per year) to low intensity continuous grazing by Sorraia horses (10 horses introduced in April 2021 and later reduced to 5 to 6 horses in 2022 and onwards) ( Supplementary Table S1 ). Site 2 has a milder and drier climate, with an average annual temperature of 13.3°C and precipitation of 544 mm ( Supplementary Figure S1 ) and includes 330 hectares of grazing area, where the grazing regime transitioned from a baseline of 20 to 40 cattle (continuous grazing) to low intensity continuous grazing by 16 Sorraia horses in 2022 ( Supplementary Table S1 ). The Sorraia horse is a native Portuguese breed adapted to living in wild or semi-wild conditions, making it well-suited for rewilding projects in landscapes that require high resilience. Its robustness allows it to endure harsh environmental conditions and survive on low-quality forage (Pinheiro et al., 2013).

Within each study area, six survey sites were selected prior to field surveys to cover representative habitats and dominant vegetation types. A paired design was implemented, where each survey site (approximately 40 m × 40 m) contained four sampling plots, each measuring 100 m² (10 m × 10 m) ( Figure 1 ). Of these, two plots were fenced as control plots to exclude grazing, allowing vegetation to naturally regenerate, and two plots were open to grazing by semi-wild horses.

Baseline data, corresponding to the existing cattle grazing regime ( Supplementary Table S1 ), were collected for all monitored variables (see next section) at Site 1 and Site 2 in 2021, before cattle removal and the introduction of horses. Subsequent monitoring of vegetation structure and aboveground biomass was conducted in 2023 and 2024, while plant composition was only reassessed in 2024.

Within each sampling plot (10 m × 10 m), vegetation structure was assessed using four perpendicular 5-meter transects originating from the plot center, spaced at 90° angles or the maximum possible angles between them. Vegetation type (grasses, forbs, and shrubs) and height class (0–0.25 m; 0.25–0.50 m; 0.5–1.3 m; 1.3–2 m; 2–4 m; >4 m) was recorded at every meter along these transects. Plant community composition was surveyed in four 1 m × 1 m quadrats, located at the end of each transect. Within each 1²m quadrat, the presence and percentage cover category of understory plant species were recorded using Braun-Blanquet’s cover scale (Kent and Coker, 1994): + (<1%); 1 (1-10%); 2 (11-25%); 3 (26-50%); 4 (51-75%), 5 (76-100%). Plants were identified to the species level whenever possible, however, some taxa were only identified at the genus level for the purpose of data analysis, to harmonize species lists and enable comparison between years ( Supplementary Tables S2, S3 ). Aboveground biomass for forbs and grasses was determined by harvesting all plant material within a 0.4 m × 0.4 m quadrat randomly placed within the sampling plot. Biomass sampling was conducted in one fenced (ungrazed) and one non-fenced (open to grazing) plot within each survey site (40 m × 40 m). Collected plant material was sorted into two functional groups (grasses and forbs), oven-dried at 60°C for 72 hours, and weighted to determine dry biomass.

Due to unforeseen land management changes in Site 1 during this study, only four of the initial six survey sites (40 m × 40 m) remained under the planned treatments. Data analysis at Site 1 was therefore based on the remaining four survey sites.

Data collected at the survey sites ( Figure 1 ) were used to characterize vegetation structure and composition, with data aggregated at the sampling plot level (10 m × 10 m). Four vegetation structure metrics were evaluated: mean grass height (cm), mean shrub height (cm), shrub cover (%) and aboveground shrub biomass (t/ha), the last metric being computed using field data on shrub cover and height ( Table 1 ). These metrics allowed the assessment of semi-wild horse grazing effects, or lack thereof, on vegetation structure in the two main vegetation layers affected by grazing, i.e., the herbaceous layer and the shrub layer, and can be used as an indicator of wildfire hazard ( Table 1 ). Vegetation structure metrics were analyzed using linear mixed models (LMMs) with separate models fitted for each site. Grazing treatment (semi-wild horse grazing vs. no grazing), year, and their interaction were set as fixed factors, and the survey site (40 m × 40 m) was specified as a random factor to account for the repeated measures over time. Assumptions of residuals’ normality and homoscedasticity were graphically checked (Zuur et al., 2009). All variables satisfied the normality and homoscedasticity assumptions. LMMs were run with the lme4 package (Bates et al., 2015).

Regarding plant community composition, three metrics were evaluated: species richness, the ratio of forbs to grasses based on biomass values, and the cumulative importance value (IV) of forbs and grasses based on species relative frequency and dominance in sampling plots ( Table 1 ). Importance values were calculated at the species level ( Supplementary Tables S2, S3 ) following Kent and Coker (1994), and then summed for forbs and grasses. Importance values were also used to identify the species with the highest IV in the baseline year and for each treatment in 2024. Forbs were defined as all species, observed at each site, excluding those from the families Poaceae, Juncaceae, Fagaceae, Dennstaedtiaceae, Aspleniaceae, and the genera Cytisus and Genista (Fabaceae). Additionally, species temporal turnover (baseline (2021) vs. last monitoring year (2024)) was assessed for each sampling plot using the codyn package (Hallett et al., 2016) to evaluate the overall compositional change in plant communities over time. Finally, the presence frequency of Stipa gigantea in 1m² quadrats was analyzed as an indicator of its dominance under each grazing treatment. This perennial grass is particularly relevant due to its fire-related traits, which contribute to fine fuel loads and fire hazard, including tall flower spikes, a dense and persistent tussock structure, and rapid post-fire regeneration (Prober et al., 2007; Cruz et al., 2017). Differences in species richness and temporal turnover were assessed through ANOVA and paired t-tests using the stats package (R Core Team, 2022). All analyses were conducted in R software (R Core Team, 2022).

In Site 1, the values of shrub cover, shrub height and shrub biomass do not suggest any differences between the effects of the grazing treatment (i.e., horse grazing or no-grazing), or survey year ( Figures 2a–c ; p > 0.05). However, mean grass height showed a significant response to the interaction between grazing treatment and survey year ( Figure 2d ). More specifically, in 2023, mean grass height increased in the ungrazed plots compared to horse grazed plots (p = 0.047), with this difference becoming more pronounced in 2024 (p < 0.001).

In Site 2, shrub cover was significantly different between the plots allocated to the different treatments at the baseline year ( Figure 3a ; p = 0.014). While not ideal, this difference was likely caused by operational constraints during the fence installation in a challenging terrain. Shrub cover showed a marginal increase between 2024 and the baseline (p = 0.051), while significant year effects were observed for shrub height and biomass, with increases in 2023 ( Figure 3b ; p = 0.003 and Figure 3c ; p = 0.024) and 2024 ( Figures 3b, c ; p < 0.001) relative to baseline. No significant year-by-treatment interactions or effects of grazing treatment alone were detected for shrub metrics. Mean grass height values ( Figure 3d ) showed significant changes related to the survey year, with an increase in 2024 (p = 0.002) relative to the baseline. The interaction between year and treatment was also significant in 2024 (p = 0.016), suggesting a more substantial increase in grass height in ungrazed plots during this year compared to horse grazed plots.

A total of 86 (2021) and 81 (2024) plant taxa were recorded across the four survey sites at Site 1 ( Supplementary Table S2 ), and 130 (2021) and 129 (2024) across the six survey sites at Site 2 ( Supplementary Table S3 ). Average richness values at the sampling plot scale were not statistically different across treatments at both Site 1 and Site 2 ( Figure 4 ; p > 0.05).

In Site 1, the forb-to-grass (F:G) ratio increased under horse grazing by 2024 ( Figure 5a ; p = 0.009; Supplementary Figure S2 ). This ratio was also marginally influenced by the year × treatment interaction in 2024 ( Figure 5a ; p = 0.056). In contrast, the F:G ratio was not statistically different across years in ungrazed plots ( Figure 5a ; p > 0.05).

In Site 2 ( Figure 5b ), the F:G ratio showed a consistent increase over time, with an overall significant year effect observed in 2024 (p = 0.015). No significant year-by-treatment interactions or effects of grazing treatment were detected ( Figure 5b ; p > 0.05; Supplementary Figure S3 ).

The analysis of the cumulative importance values (IV) of grasses and forbs ( Table 2 ) indicates similar patterns of community change at both Site 1 and Site 2. Namely, a more pronounced decline in the cumulative IV of grasses and a greater increase in that of forbs from baseline to horse grazing conditions. In terms of dominant species (i.e., those with the highest IV), composition remained relatively stable across years and treatments at both sites ( Table 2 ), although some notable shifts were observed. At Site 1, Cytisus multiflorus and Stipa gigantea increased their IVs, and Avena barbata (a tall grass species) emerged as a new dominant in ungrazed plots. At Site 2, C. multiflorus not only consistently exhibited the highest IV but also increased in dominance over time, with treatment differences appearing less influential.

Species turnover in Site 2 was significantly higher in the fenced plots, i.e., that shifted from the baseline to an ungrazed regime ( Figure 6b ; p = 0.045), whereas differences were not statistically significant at Site 1 ( Figure 6a ; p > 0.05).

The presence frequency of S. gigantea in 1 m² plots increased in the ungrazed areas at Site 1 (20 out of 32 plots in 2024 vs. 15 out of 32 in 2021) but remained stable in the horse grazed areas (13 out of 32 in 2024 vs. 12 out of 32 in 2021). In Site 2, the species was not recorded in any plot in 2021 but was observed in one fenced 10 × 10 m sampling plot in 2024.

The findings for vegetation structure partially align with our first hypothesis (H1), with horse grazing showing a clear effect on herbaceous vegetation but not on shrubs. The lack of significant treatment effects (i.e., grazed vs. ungrazed plots) on shrub cover and height at both sites ( Figures 2 , 3 ) suggests that horse grazing may not exert sufficient pressure to regulate woody vegetation. This is consistent with previous studies indicating that low density grazing by domestic herbivores, particularly horses, may not suppress shrub expansion in abandoned areas of the Mediterranean region (Moinardeau et al., 2016; Fagúndez et al., 2022; Ribeiro et al., 2023, 2024). Moreover, the steeper increase in shrub metrics observed over the years at Site 2 may be partially driven by post‐fire successional processes, as this site recently experienced a severe wildfire in 2017. Fire can reduce competing vegetation, create open space, and release nutrients, thereby promoting rapid regrowth of fire‐adapted shrub species (Cruz et al., 2020; Magaña Ugarte et al., 2021; Alegria, 2022). These favorable post‐disturbance conditions may enable shrub expansion under low grazing pressure (Bates and Davies, 2014; Smit and Coetsee, 2019; Siegel et al., 2022).

The similar trends observed for grass height at both sites, with the significant interaction between year and treatment, indicate that even in the short term, horse grazing can effectively limit grass height, compared to grazing exclusion. These findings reinforce the role of horses as herbaceous grazers, preferentially consuming grasses over woody species (Vulink et al., 2001), whereas cattle tend to include a higher proportion of woody plants in their diet (Cosyns et al., 2001; Menard et al., 2002; Lamoot et al., 2005). The increased presence of the fire-prone Stipa gigantea in ungrazed (control) plots in Site 1 suggests that the lack of grazing pressure may favor its establishment and aligns with previous research indicating that grazing exclusion can promote grass dominance (de Villalobos and Zalba, 2010; Schneider and Hering, 2024), potentially elevating fire hazard in these areas (Cardoso et al., 2022; Davies et al., 2022; Orr et al., 2023). Conversely, the stability of Stipa gigantea’s occurrence in open areas indicates that horse grazing may help contain its spread.

In terms of vegetation composition, although we expected a decline in species richness in ungrazed (control) plots (H2) (Papanikolaou et al., 2011; Henning et al., 2017a; Dvorský et al., 2022; Bonavent et al., 2023), no significant differences were observed in species richness between grazing treatments (horses vs. no grazing) at either site. Despite the lack of significant effects, Figure 4 shows a trend towards lower species richness in ungrazed plots, with decreasing median and mean, while plots that remained grazed maintained similar richness levels. One potential explanation for the absence of statistically significant effects may be the early monitoring stage, as the short duration of grazing exclusion in fenced plots may not yet be sufficient to produce detectable effects (Song et al., 2020). Additionally, at Site 2, species temporal turnover was significantly higher in ungrazed (control) plots when compared to horse-grazed (open) plots, while no significant differences were detected at Site 1. These results provide partial support for our third hypothesis (H3), indicating a greater shift in species composition over time in the absence of grazing. It also suggests that the removal of grazing may lead to shifts in the recruitment success among plant species, such as grazing adapted species that may decline in response to exclusion (Song et al., 2020) or competitively dominant species that were previously suppressed by grazing (Zhang et al., 2023). This result aligns with other studies showing that even when species richness remains unchanged, the cessation of grazing can lead to significant changes in plant community composition over time, as certain species may outcompete others (Schultz et al., 2011; Bar-Massada and Hadar, 2017; Kaufmann et al., 2021).

Furthermore, as expected in our fourth hypothesis (H4), which anticipated a shift in the relative representation of forbs and grasses under different grazing pressure, there was an observed increase in forb-to-grass (F:G) ratio under horse grazing at Site 1, providing additional insights into how this grazing regime can interact with plant functional groups. In contrast, ungrazed plots showed no significant changes in the F:G ratio ( Figure 5 ), suggesting that forb recruitment may be limited without grazing disturbances and may benefit from reduced competition for space with grasses (Fleurance et al., 2012; Tuomi et al., 2019). These results might also help explain the results observed in species temporal turnover, as changes in turnover can be influenced by changes in abundance (Peper et al., 2011; Hillebrand et al., 2018) and grasses are outcompeting forbs in ungrazed plots, potentially leading to lower structural and functional diversity (Yan and Liu, 2021; Cardoso et al., 2022). Additionally, these functional responses align with the results obtained by the analysis of species importance values (IV), which revealed parallel trends in community structure. At both sites, the sum of IV for grasses declined more sharply in horse-grazed plots compared to ungrazed areas, while forbs increased their cumulative IV under grazing. Both metrics suggest that natural horse grazing promotes a shift toward more forb dominated communities (with flowering plants), which enhance food and shelter resources for anthophilous insects (Lázaro et al., 2016; Henning et al., 2017b; Shapira et al., 2020; Dvorský et al., 2022; Garrido et al., 2022; Bonavent et al., 2023), while grazing exclusion favors grass dominance ( Supplementary Figures S2 , S3 ). Notably, Stipa gigantea increased its importance in ungrazed plots, reinforcing the previously discussed concerns about fine fuel accumulation (Menning and Stephens, 2007; Cardoso et al., 2022; Davies et al., 2022; Orr et al., 2023). Grass dominance in the absence of grazing could have adverse implications for wildfire hazard and biodiversity conservation, particularly in fire-prone Mediterranean landscapes (Kirkland et al., 2024; Lovreglio et al., 2024), such as the Côa Valley region.

This study provides relevant evidence on the use of semi-wild horse grazing as a tool for vegetation management. On the one hand, horse grazing demonstrated an ability to reduce local fire hazard by effectively controlling the height of grasses ( Figures 2 , 3 ), while also promoting a higher representation of forbs in the plant community, as reflected by increases in the forb-to-grass ratio and cumulative importance values ( Figure 5 and Table 2 ). On the other hand, the limited effects on shrub metrics at both sites, with increased shrub cover, height and biomass ( Figures 2 , 3 ), suggest that horse grazing alone may be insufficient to regulate woody vegetation. Furthermore, preliminary results obtained from soil and tree recruitment analysis developed in a management action for the study sites showed that shifting the grazing regimes improved soil organic matter (SOM) and oak recruitment ( Supplementary Table S4 and Supplementary Figure S4 ), with greater increases in SOM and higher number of oak seedlings in horse grazing plots, compared to ungrazed plots. Overall, these findings suggest that while semi-wild horse grazing can contribute to multiple management goals, its effectiveness varies across vegetation and habitat types and would benefit from complementary interventions tailored to local vegetation dynamics and disturbance history.

In particular, semi-wild horse grazing alone may not be sufficient to achieve the desired management outcomes for shrub control in the Côa Valley region, which has been affected by recurrent and severe wildfires over the last two decades (Kirkland et al., 2024). The unchecked proliferation of tall grasses can also significantly heighten fire hazard (Menning and Stephens, 2007; Cardoso et al., 2022), which is particularly concerning under the hot and dry conditions typical of Mediterranean summers. However, in grassland-like habitats, where the accumulation of fine fuels can enable rapid fire spread (Turco et al., 2017; El Garroussi et al., 2024), low-intensity semi-wild horse grazing can be effective in preventing grass dominance and reducing wildfire risk, without imposing the grazing pressure that is needed to suppress shrub expansion. Notably, this finding signals the important ecological role of diverse and complementary herbivore communities (Orr et al., 2022; Pringle et al., 2023; Ribeiro et al., 2023). Management strategies should incorporate a diverse assemblage of herbivores, as this is likely to create a mosaic of grazing effects, fostering ecosystem function and resilience in Mediterranean landscapes (Liu et al., 2015; Orr et al., 2022). For example, while horses may be well-suited to certain contexts due to their lighter impact and broad grazing patterns, locally adapted cattle breeds and other rustic breeds, such as Tauros, can complement these effects by targeting different vegetation types or structures (Fleurance et al., 2012; Moinardeau et al., 2016; Schmitz and Isselstein, 2020). This diversified approach is particularly important in habitats prone to shrub encroachment, where it can help break fuel continuity and maintain landscape heterogeneity (Pausas, 2004; Nunes, 2023). Additionally, rewilding initiatives, where herds are semi-wild, social, and not reliant on supplementary feeding (except under extreme conditions), grazing patterns tend to be more heterogeneous than those in livestock production systems (Menard et al., 2002; López et al., 2019). This heterogeneity can promote a more dynamic and resilient vegetation structure by preventing the dominance of single plant functional groups (Adler et al., 2001; Pringle et al., 2023). An important consideration in implementing such management strategies is the definition of appropriate stocking densities for different herbivore assemblages. However, determining adequate densities is challenging, as grazing pressure is not defined by stocking density alone, and its ecological effects are highly context-dependent (Ribeiro et al., 2023). Moreover, in addition to promoting a diverse assemblage of herbivore, complementary management actions, such as initial shrub clearing and prescribed burning, might be needed to support vegetation regulation efforts. These actions can be particularly relevant in shrub encroached areas (Castro et al., 2022; Oikonomou et al., 2023) or where prescribed burning can reduce accumulated litter and biomass and foster herbaceous growth (Barbaro et al., 2001; O’Connor et al., 2020).

In the specific context of the study sites, and of ongoing rewilding efforts, Rewilding Portugal plans to introduce Taurus cattle (Goderie et al., 2013; Stokstad, 2015) at lower densities than those of the pre-existing commercial livestock production, aiming to complement the effects of horse grazing. As this rewilding plan is implemented, ongoing monitoring will be necessary to assess its impacts and provide evidence to evaluate this approach. Moreover, the significant effects of the year factor across sites highlight the need to account for temporal variability in future grazing management plans, as vegetation responses may lag behind grazing interventions (Song et al., 2020; Kaufmann et al., 2021; Ribeiro et al., 2023). Additionally, the interannual variation in weather conditions may have influenced plant growth and community dynamics, potentially interacting with the grazing treatments. Disentangling the effects of grazing and climatic variability also requires a longer monitoring interval to better understand the relative contribution of these factors to the observed changes (Kutiel et al., 2000; Kaufmann et al., 2021). Despite these limitations, early-stage assessments remain essential to detect initial vegetation responses (e.g. Balata et al., 2022; Li and Zhan, 2023), which can guide the adaptive management of rewilding initiatives, especially in fire-prone landscapes where early signs of structural change can inform hazard mitigation.